Walrus Memory release notes

MCP

Walrus Memory MCP v0.0.16

October 26, 2026 | GitHub

This release updates @modelcontextprotocol/sdk to 1.32.1, which includes transport security fixes.

Authenticated SSE GET and POST requests now refuse redirects. The message endpoint that the server supplies must stay on the configured relayer origin. Because of this, custom authentication headers and tool arguments cannot be forwarded to a redirected origin. If your deployment relies on redirects to reach the relayer, configure the final relayer URL directly.

---

Walrus Memory MCP v0.0.15

October 21, 2026 | GitHub

A project-local .memwal/credentials.json is ignored until you approve that exact project path, account, delegate key, and relayer with memwal-mcp approve-project. Approvals are stored in ~/.memwal/project-approvals.json, and any change to the destination requires approving again. Until then, the global credentials are used and a stderr line names the skipped file. memwal_health now reports account= beside relayer=.

MEMWAL_CREDS_DIR must now be an absolute path outside the current project. Other values are refused with an error, and an empty value means unset. The login write-ahead record moves out of the repository to ~/.memwal/login-pending/. memwal_login with freshKey: true now closes the previous listener and names the abandoned public key and address so you can remove them from the dashboard.

---

Walrus Memory MCP v0.0.14

September 22, 2026 | GitHub

The plugin no longer starts the MCP server through npx, which could run a package planted in the open project. Every launch site now runs plugin/scripts/launch_mcp.mjs. It installs the pinned version under ~/.memwal/runtime and fails rather than falling back. MEMWAL_MCP_RUNTIME_DIR relocates that directory and must be an absolute path.

memwal_remember and memwal_remember_bulk now return a job_id at accept by default. Set MEMWAL_MCP_REMEMBER_WAIT_MS=90000 to wait for a blob_id. Set MEMWAL_MCP_TRANSPORT=http to use Streamable HTTP, though replayed writes might duplicate. Timeouts report the cause and relayer health. The bridge honors retry_after, memwal_remember_status accepts job_ids, and delegate keypairs persist before browser sign-in.

Changes

• The cold-start tool list no longer advertises a tool the relayer may not serve. The bridge ships on npm and updates itself while a relayer ships per environment, so 0.0.14-dev.0 dialled prod and staging still on 0.0.13: cold start named memwal_remember_status, which neither registers, and the pending-write wording sent the agent to go call it — one live run spent 90.67s there before erroring. Cold start is now a floor rather than a forecast (BASELINE_RELAYER_TOOLS): it carries only what the oldest supported relayer serves and its descriptions name nothing outside it, while newer tools still reach the client a beat later on the relayer's own tools/list. Initialize instructions use the same floor (they used to name memwal_remember_status before any tools/list). And a call for a tool outside that floor — including during the cold-start window, before any upstream tools/list has been seen — is now answered locally and at once — naming the tools that do exist and saying plainly that nothing ran — instead of being forwarded into a wait that only ends at the orphan deadline.

---

Walrus Memory MCP v0.0.13

September 15, 2026 | GitHub

Timed-out writes from memwal_remember, memwal_remember_bulk, and memwal_analyze now warn that the write might have completed and tell you to check with memwal_recall before re-saving, which avoids duplicate paid copies. The bridge reports a rejected delegate key immediately and points to memwal_login. It backs off on HTTP 429 using Retry-After. It fails requests stuck behind a failing handshake after 90s, which you can override with MEMWAL_MCP_STALLED_HANDSHAKE_MS.

After an in-session memwal_login, the bridge keeps reading stdin and confirms the sign-in. The CLI writes credentials as a new 0600 file, warns about unrecognized options, and lists network presets in --help. memwal_health now reports the relayer URL, and memwal_restore retries a page after a transient failure.

Changes

• Unrecognised options now warn on stderr and in the structured log (cli.unrecognised_arg) instead of being dropped in silence, so a typo'd --namesapce work no longer writes to the default namespace with nothing to say it had. The warning names the option key only, keeping a mistyped value-taking flag (--tokenn=hunter2) from putting the secret on stderr, and it warns rather than exits so an option from a newer config cannot brick the server.
• --help now lists the network presets (--prod, --dev, --staging, --local) with the relayer and web URLs each resolves to, rendered from the preset table rather than retyped so a new preset cannot ship undocumented the way --prod did. The --label default is corrected to "MCP Client", which is what the code actually falls back to.
• memwal_health reports relayer=<url>, naming the origin this process dialled, so a client bound to the wrong network finds out there instead of by noticing its memories are missing. The URL is captured when the call is sent rather than when the reply lands, so a reconnect mid-flight cannot label the answer with a relayer it did not come from.
• Keep reading stdin after an in-session memwal_login. The auth-required stub hands off to the bridge by pausing stdin, and a stream paused that way does not resume when a new data listener attaches, so the bridge served only the request replayed from the hand-off and then went deaf. Signing in appeared to work, the retry succeeded, and every call after it hung until the client timed it out.
• Confirm a completed sign-in instead of only writing it to the log file. The failure path already reported itself twice (a notification and a notice on the next tool call) while success reported nothing, so a user who approved in the browser could not tell whether credentials had landed, the bridge had adopted them, or a retry was worth trying. Success now sends the matching notification and prefixes a one-shot banner naming the account, the delegate, and the resolved credentials path onto the next tool result.
• Stop the memwal_login prompt claiming you are already signed in when you are not. The bridge assumed it only ever runs with credentials on disk, but memwal_logout deletes them and login is intercepted before the signed-out guard, so a login after a logout in the same session announced that a stored delegate key would be replaced, reading as though the logout had not taken. Both prompts now read the credentials file instead of assuming the mode.
• Write the credentials file by creating a new 0600 file and renaming it into place, instead of writing the delegate private key to the existing path and tightening the permission afterwards. writeFileSync's mode only applies when it creates the file, so a credentials.json left world-readable by anything outside the CLI (a manual chmod, a restored backup, another tool) received the new key under the old permission until the following chmod landed. The CLI now writes the backup it takes when signing in as a different account the same way; that backup holds the same plaintext key, and the CLI previously created it under the process umask. On Windows, where rename cannot replace a destination another process holds open and POSIX mode bits are not enforced at all, the CLI retries briefly and then writes in place rather than failing the sign-in.

---

Walrus Memory MCP v0.0.12

September 9, 2026 | GitHub

memwal_recall accepts an optional maxDistance cosine-distance cutoff, which must be 0 or greater. It drops hits with distance >= maxDistance, and result lines now show both score and distance. The server forwards the client's initialize.clientInfo to the relayer as x-memwal-client and x-memwal-client-version, so sidecar logs name the coding agent.

The credential directory now resolves on every access, and MEMWAL_CREDS_DIR overrides it. This fixes a Windows problem where the test suite overwrote ~/.memwal/credentials.json and its delegate key. memwal_restore now clarifies that truncated=true is retryable and that truncated=false does not guarantee completeness.

Changes

• Optional maxDistance on memwal_recall: cosine-distance cutoff (low = similar, must be >= 0). Hits with distance >= maxDistance are dropped. Result lines now include both score (1 − cosine distance) and distance.
• Resolve the credential directory on every access instead of freezing it at module load, and let MEMWAL_CREDS_DIR override it. The login test sandboxed the home directory with HOME alone, which os.homedir() ignores on Windows, so running the package's test suite there wrote fixture credentials over the developer's real ~/.memwal/credentials.json and destroyed the delegate key stored in it.

---

Walrus Memory MCP v0.0.11

August 24, 2026 | GitHub

This release fixes MCP bridge reliability. The SSE handshake socket no longer aborts on HTTP 429, 401, or a bad content-type, which crashed on Windows. Outbound MCP POSTs are serialized, and the bridge reconnects when the SSE handle is gone. A burst of memwal_recall calls no longer breaks it until restart.

login without a TTY exits with code 1 instead of reporting success. Concurrent memwal_login calls reuse the in-flight listener and URL rather than starting a second flow that can hang later remember and recall calls. A login timeout warns that an on-chain delegate key might already exist if the browser approved after the listener expired. Session occupancy caps are reported as concurrent limits rather than a 30-second cooldown.

---

Walrus Memory MCP v0.0.10

August 20, 2026 | GitHub

The MCP server now sends proactive-usage instructions in the MCP initialize handshake. The model knows when to save and recall memories even when your client lazily loads tool descriptions. The bridge carries these instructions at cold start. Signed-out sessions get guidance pointing at memwal_login and a conservative tools/list. memwal_recall is advertised as read-only. The UserPromptSubmit hook lets the agent choose between remember and recall from meaning rather than keywords.

Tool calls whose upstream response never arrives return a retryable error after MEMWAL_MCP_CALL_TIMEOUT_MS (default 240s) instead of hanging. memwal_remember_bulk uses your configured namespace (--namespace / MEMWAL_NAMESPACE) when a call omits one. serverInfo.version reports the real package version.

Changes

• Answer orphaned tool calls whose upstream response never arrives with a retryable error instead of hanging indefinitely. The bridge now tracks in-flight request start times and sweeps expired calls through a per-request deadline (MEMWAL_MCP_CALL_TIMEOUT_MS, default 240s), reuses the existing late-reply drop so an expired call cannot get a second response, and enriches reconnect logs with pending request IDs and methods.
• Inject the configured default namespace (--namespace / MEMWAL_NAMESPACE) into memwal_remember_bulk calls that omit one, so bulk facts land in the project namespace instead of the relayer fallback default.
• Send proactive-usage instructions in the MCP initialize handshake, so the model knows when to save and recall without being asked. Clients moved to lazy tool loading, which keeps tool descriptions out of the model's context until a tool is explicitly loaded; the guidance lived only in those descriptions, so the model stopped using memory on its own and would offer its built-in memory or deny the tool existed. instructions travels with initialize, before any tools/list, so lazy loading cannot strip it.

---

Walrus Memory MCP v0.0.8

August 14, 2026 | GitHub

The Walrus Memory MCP server now includes human-readable titles for its tools during pre-login tool discovery. Each tool also carries explicit readOnlyHint and destructiveHint metadata at that stage. This metadata matches what the remote relayer provides to Claude connectors, so clients that inspect tools before login get the same titles and hints as they would from the relayer.

---

Walrus Memory MCP v0.0.7

August 14, 2026 | GitHub

This release contains three fixes. After you log in through the browser, the MCP server now reloads your credentials, so you no longer need to restart the MCP client.

Handshake handling is stricter. The server rejects stale handshakes and prevents requests from being replayed across account changes.

Restores now report when results are truncated, and the server enforces the relayer's bounded restore limit.

---

Walrus Memory MCP v0.0.6

July 31, 2026 | GitHub

This release tightens the delegate-key login flow. Before it accepts a login callback, the MCP server now requires a localhost preflight handshake that proves the exact state, public key, and relayer. The server also takes the canonical login label from the verified local flow and no longer trusts the label that the browser callback supplies.

MCP SSE POST messages and replayed requests now authenticate with the delegate credentials.

---

Package v0.0.5

June 12, 2026 | GitHub

Adds the automatic memory plugin for Claude Code, Codex, Cursor, and Antigravity, along with new memwal_remember_bulk and memwal_health tools and proactive memory behavior. Also fixes the plugin bundle so it ships .mcp.json correctly.

---

Package v0.0.4

June 5, 2026 | GitHub

Fixes HTTPS dashboard sign-in callbacks and credential reload after login so tools work without restarting the MCP client.

---

Package v0.0.3

June 4, 2026 | GitHub

Rebrands package metadata from MemWal to Walrus Memory throughout the MCP server.

---

Package v0.0.2

May 25, 2026 | GitHub

Adds relayer compatibility checks and rebrands from MemWal to Walrus Memory across the MCP server package.

---

Package v0.0.1

May 15, 2026 | GitHub

Initial release of the MCP server with stdio transport, browser wallet login, and inline session tools. Includes memory tools for remember, recall, analyze, and restore.

---

TypeScript SDK

Walrus Memory TypeScript SDK v0.1.10

October 26, 2026 | GitHub

This release tightens input and response validation in the Walrus Memory SDK. recallManual on MemWal and MemWalManual rejects NaN or Infinity in scoringWeights before signing, matching recall. Previously, JSON.stringify sent these values to the relayer as null. remember, rememberBulk, analyze, embed, and recall also reject text containing only whitespace and zero-width characters. Previously, these calls waited until the client timed out. MemWalMock rejects the same input.

When a 2xx response has the wrong JSON shape, the SDK raises an error naming the endpoint and the missing array instead of a raw TypeError. The SDK does not retry these bulk status errors. A null or non-object status body makes getRememberStatus return not_found and waitForRememberJob throw MemWalRememberJobNotFoundError.

---

Walrus Memory TypeScript SDK v0.1.9

October 21, 2026 | GitHub

createAccount, addDelegateKey, removeDelegateKey, and MemWalManual work again without a suiClient. The default is now SuiGrpcClient against the public fullnode for suiNetwork. rememberBulk and rememberBulkAsync send a derived idempotency key, so a retry within 30 minutes returns the same jobs. listMemories() pages stored memories as metadata without a similarity query.

Remember-job waits back off toward a 5s cap. On timeout they raise MemWalRememberJobTimeout with lastStatus and serverError, and they raise MemWalRateLimited when status reads stay rate-limited. Many methods reject invalid input before signing, including bad limit, maxDistance, job ids, and blank text. Package subpaths now resolve in Next.js, Turbopack, and Node require(esm), and /manual no longer imports @mysten/walrus.

Changes

• rememberBulk and rememberBulkAsync send an idempotency key derived from the batch (a 30-minute bucket plus each item's namespace and text). A retry inside that window returns the same jobs instead of storing the batch again. An explicit idempotencyKey still wins.
• withMemWal does not recall or auto-save on a tool-execution step. A later tool step reuses the recall cached for that user message. Auto-save still runs only on a user turn.
• flush() raises when a background auto-save failed, and onSaveError runs. The chat turn itself still does not fail.
• waitForRememberJob and waitForRememberJobs do not sleep past the deadline. If the budget runs out during the sleep, they make one more status read.
• remember, rememberAsync, and rememberAndWait reject a non-string text argument with TypeError before signing. An object used to be sent as text and the relayer answered 422.
• MemWalMock.rememberBulkAsync rejects an empty or non-array batch with the same error as MemWal.
• MemWalMock honors idempotencyKey. A retry of the same text and namespace returns the original job, a different payload throws, and forget / clear release the key so the text can be written again.
• MemWalMock.recall({ sort: "recent" }) orders a bounded semantic window (5× limit, at least limit, at most 50) by write time before truncating. Relevance order is unchanged.
• Positional recall(query, options, namespace) uses that namespace when options.namespace is omitted, on both MemWal and MemWalMock. An explicit options.namespace still wins.
• remember(text, { idempotencyKey }) and rememberAndWait(text, { timeoutMs }, { idempotencyKey }) treat a plain object in the namespace slot as options. A third options argument is merged, and its fields win. That call used to send the object as the namespace, or drop the third argument and mint a second blob. An explicit string namespace still wins.
• getRememberStatus("") returns not_found without a request, and waitForRememberJob("") fails immediately with status 404. GET /api/remember/ is not a route; the empty 404 body used to throw SyntaxError, and the wait treated that as transient until timeout.
• rememberBulkAsync rejects more than 20 items before signing, matching the relayer's MAX_BULK_ITEMS.
• Remember-job status reads (getRememberStatus, waitForRememberJob, getRememberBulkStatus) no longer attach x-seal-session. Status is metadata and does not decrypt.
• Bulk status includes each job's namespace. waitForRememberJobs(jobIds) uses that value when the caller did not pass one, instead of the client default. A caller-supplied namespace still wins.
• getPublicKeyHex() and any request after destroy() throw, including a read already in flight. getPublicKeyHex() used to resolve to the zeroed key, or to the well-known key of the wiped seed, and remember() still sent GET /version before noticing the client was destroyed.
• normalizeServerUrl no longer warns about plaintext HTTP for IPv6 loopback. URL.hostname keeps the brackets on [::1], so the ::1 exemption never matched http://[::1]:8000.
• remember, rememberBulk, analyze, and embed reject whitespace-only text immediately, with the same messages the relayer returns. A blank body used to be accepted and then sit until the client timed out.
• A 401 AUTH_REJECTED now says a wallet address is not a MemWal account ID. The relayer still returns one 401 for every credential failure, so the message names the mix-up instead of revealing which check failed.

---

Walrus Memory TypeScript SDK v0.1.8

September 22, 2026 | GitHub

All relayer requests now carry a deadline, so one stalled request no longer runs past timeoutMs. Requests default to 30 seconds, and you can change this with requestTimeoutMs on MemWal.create. restore and analyze each use 60 seconds. An expired request raises MemWalRequestTimeout with status: 504, which the existing poll handling retries.

Generated idempotency keys now derive from the namespace, the text, and a 30-minute bucket. A replay from a new client therefore no longer creates a duplicate paid Walrus blob. Explicit idempotencyKey values are unaffected. recall() also sends deadline_ms. Relayers that read it return a 504 with serverCode RECALL_TIMEOUT that names the stuck step.

---

Walrus Memory TypeScript SDK v0.1.7

September 15, 2026 | GitHub

restore() results now include a failed count for permanent decrypt or UTF-8 failures. These were previously folded into skipped or dropped silently. The field is required, like truncated, and the SDK defaults it to 0 when omitted.

Request body hashing now uses @noble/hashes instead of WebCrypto with a node:crypto fallback. Browser apps built with Vite no longer crash on the first remember or recall. The package now declares engines.node >= 20.0.0. Empty-body 401 responses return the AUTH_REJECTED troubleshooting message instead of suggesting memwal_login. Transactions in account.ts and manual.ts now use typed tx.pure helpers that work with modern @mysten/sui.

---

Walrus Memory TypeScript SDK v0.1.6

September 9, 2026 | GitHub

recall() results include an optional created_at field with the RFC3339 write-time of the stored fact. RecallOptions accepts sort: "relevance" | "recent" and scoringWeights. With sort: "recent", the SDK over-fetches semantic candidates (5x limit, capped at 50), orders them by write-time descending, and then truncates the results to limit.

The SDK now reports an HTTP 503 with x-auth-error: AUTH_UPSTREAM_UNAVAILABLE as a retryable credential-verification outage instead of a sign-in failure. Other 503 responses keep the generic sanitized body, and empty-body 401 responses still point you to memwal_login.

---

Walrus Memory TypeScript SDK v0.1.5

August 26, 2026 | GitHub

rememberManual now sends encryptedData, a base64 SEAL ciphertext, instead of a pre-uploaded blobId, matching POST /api/remember/manual. recall() results include dropped_count when the relayer omits matches that fail to download or decrypt. health() surfaces an optional write_ready field when the relayer reports write-path sidecar liveness.

Manual recall now decodes SEAL object IDs with hexToBytes. This removes the risk of coercing NaN to 0 or throwing on empty input. HTTP error messages and remember-job error strings no longer expose localhost or 127.0.0.1 URLs to callers.

---

Walrus Memory TypeScript SDK v0.1.4

August 24, 2026 | GitHub

This release improves the error message for empty-body 401 responses. If you are calling through MCP for the first time and receive an empty-body 401, the SDK now tells you to run memwal_login. Previously it showed <no message>, which gave no indication of what to do next.

Credential 401s keep their existing behavior and still point you to the AUTH_REJECTED troubleshooting guide.

---

Walrus Memory TypeScript SDK v0.1.3

August 20, 2026 | GitHub

The SDK now estimates tokens on the client side and supports token budgets for recall. When recalled content exceeds a budget, you can choose how the SDK truncates it by setting one of the configurable truncation strategies.

This release also changes the default SDK relayer URL to https://relayer.memory.walrus.xyz. If your application relies on the default relayer rather than setting one explicitly, it uses the new URL after you upgrade.

---

Walrus Memory TypeScript SDK v0.1.2

August 14, 2026 | GitHub

This release adds deterministic offline mock clients, so you can run tests without credentials. It also adds idempotency keys, which collapse retries onto a single paid remember job.

Fixes isolate foreign manual-recall blobs and safely encode large encrypted payloads. The SDK now supports both Zod 3 and Zod 4 as peer dependencies.

---

Walrus Memory TypeScript SDK v0.1.1

August 7, 2026 | GitHub

This release adds flush(), which you can call to handle pending auto-saves from withMemWal. Restore results now include a truncated field that indicates when a recovery is incomplete, so your code can detect partial restores.

AUTH_REJECTED errors now link to troubleshooting guidance, giving you a direct path to resolving authentication rejections.

---

Walrus Memory TypeScript SDK v0.1.0

July 31, 2026 | GitHub

Recalled AI memory now arrives as nonce-delimited, explicitly untrusted data instead of content injected directly into the system role.

This release adds sponsor authorization helpers for sponsored transaction flows. It also adds migration-aware account and manual configuration, including registry support and support for current Sui transaction results. Account transaction result handling and failure diagnostics improve across both legacy and current Sui response shapes.

---

Package v0.0.7

June 2, 2026 | GitHub

Adds an optional occurredAt parameter to analyze() for temporal anchoring of extracted facts. Relative date references are now resolved into absolute dates.

Changes

• Add optional occurredAt to analyze() and analyzeAndWait() for temporal anchoring of extracted facts.

---

Package v0.0.6

May 28, 2026 | GitHub

Adds RecallParams for object-style recall calls and deprecates the positional recall overload. Also documents restore() response fields.

---

Package v0.0.5

May 25, 2026 | GitHub

Adds relayer compatibility checks, RecallOptions for topK/namespace/maxDistance filtering, and prefers Sui gRPC for SEAL sessions. This release also rebrands from MemWal to Walrus Memory.

---

Package v0.0.4

May 15, 2026 | GitHub

Adds getRememberStatus for polling async job state and SealServerConfig for manual SEAL configuration. Manual mode validates API key pairs.

---

Package v0.0.3

May 8, 2026 | GitHub

Updates remember() for the async relayer flow and introduces rememberAsync, waitForRememberJob, rememberAndWait, and bulk remember helpers. Also adds async analysis with analyzeAndWait.

---

Package v0.0.2

April 30, 2026 | GitHub

Adds per-request nonce signing to block replay attacks and replaces delegate key transport with ephemeral SEAL sessions. SDK versions that do not include a nonce are rejected with HTTP 426.

---

Python SDK

Walrus Memory Python SDK v0.2.1

October 26, 2026 | GitHub

A maintenance release with no user-facing changes documented in the release notes; see the full commit log for internal updates.

---

Walrus Memory Python SDK v0.2.0

October 23, 2026 | GitHub

A maintenance release with no user-facing changes documented in the release notes; see the full commit log for internal updates.

---

Walrus Memory Python SDK v0.1.12

October 21, 2026 | GitHub

A maintenance release with no user-facing changes documented in the release notes; see the full commit log for internal updates.

---

Walrus Memory Python SDK v0.1.11

September 22, 2026 | GitHub

A maintenance release with no user-facing changes documented in the release notes; see the full commit log for internal updates.

---

Walrus Memory Python SDK v0.1.10

September 15, 2026 | GitHub

A maintenance release with no user-facing changes documented in the release notes; see the full commit log for internal updates.

---

Walrus Memory Python SDK v0.1.9

September 9, 2026 | GitHub

A maintenance release with no user-facing changes documented in the release notes; see the full commit log for internal updates.

---

Walrus Memory Python SDK v0.1.8

August 26, 2026 | GitHub

A maintenance release with no user-facing changes documented in the release notes; see the full commit log for internal updates.

---

Walrus Memory Python SDK v0.1.7

August 14, 2026 | GitHub

A maintenance release with no user-facing changes documented in the release notes; see the full commit log for internal updates.

---

Walrus Memory Python SDK v0.1.6

August 7, 2026 | GitHub

A maintenance release with no user-facing changes documented in the release notes; see the full commit log for internal updates.

---

Walrus Memory Python SDK v0.1.5

July 31, 2026 | GitHub

A maintenance release with no user-facing changes documented in the release notes; see the full commit log for internal updates.

---

Package v0.1.4

June 2, 2026 | GitHub

A patch release for the Walrus Memory Python SDK.

---

Package v0.1.3

May 28, 2026 | GitHub

A patch release for the Walrus Memory Python SDK.

---

Package v0.1.2

May 25, 2026 | GitHub

A patch release for the Walrus Memory Python SDK.

---

Package v0.1.0

May 21, 2026 | GitHub

Initial release of the Walrus Memory Python SDK.

---

OpenClaw

Walrus Memory OpenClaw v0.0.7

October 2, 2026 | GitHub

memory_search and memory_store now operate only within the calling agent's namespace. If you omit the namespace, the tools use that agent's namespace rather than the main namespace. A namespace that the model supplies is rejected unless it matches the agent's namespace or its legacy namespace.

Configuration interpolation of ${ENV_VAR} now applies only to privateKey. All other fields, including serverUrl, are read literally, so a secret cannot end up in serverUrl.

---

Walrus Memory OpenClaw v0.0.6

August 20, 2026 | GitHub

openclaw plugins install now succeeds, and the plugin declares contracts.tools, so memory_search and memory_store register. Config validation happens at register time with per-field errors and leaves the gateway running. Every relayer call now has a deadline, set with requestTimeoutMs (default 10000, range 1000 to 60000). A hung relayer no longer blocks the agent turn.

The prompt-injection filter no longer silently drops ordinary statements such as "I run the deploy command every Friday." The docs now cover hooks.allowConversationAccess, which auto-capture requires. They also correct the config key: both plugins.slots.memory and the plugins.entries key use memory-memwal.

---

Package v0.0.5

June 15, 2026 | GitHub

Fixes a workspace dependency that broke npm installs outside the monorepo and switches the release workflow to pnpm publish.

---

Package v0.0.4

June 2, 2026 | GitHub

Adds temporal anchoring through occurredAt to the memory_store tool. The auto-capture hook now passes the current timestamp, and the SDK is bumped for the new AnalyzeOptions.

Changes

• Wire temporal anchoring through the agent-side memory tools.

---

Package v0.0.3

May 25, 2026 | GitHub

Rebrands from MemWal to Walrus Memory across the OpenClaw memory plugin.

---

Package v0.0.2

April 30, 2026 | GitHub

Initial release of the OpenClaw memory plugin with automatic recall and capture hooks, session summary, CLI commands, and LLM tools.

---