Open Agent Memory v0.0.2

Open Agent Memory 0.0.2

Base: c0f2d2e

Scope: c0b29701921a366aefafc13cf9bfc4fac6308336e83a9d1e0fa1d31bb0e96311

Changes

Added

• List active users with administrator-only GET /v1/users, offset pagination and an inclusive last_write_at_gte filter. The activity timestamp tracks accepted interaction, Resource and Profile Data revisions; it is separate from indexing readiness and human-only conversation activity.
• Display stable, human-readable Insight titles in memory inspection, portable exports and Memory Inspector. Title updates preserve complete supported text, retrieval labels and provenance; legacy records have migration/read fallbacks.
• Connect an existing OAM deployment through the thin local stdio MCP server, with six user-scoped tools, bounded HTTP responses and documented Claude Code/Cursor setup. MCP remains independently versioned at 0.0.1.
• Provide multilingual onboarding, a cited-recall quickstart, published historical benchmark documentation and versioned release notes with complete documentation-review checks.

Fixed

• Keep directly retrieved lexical user originals eligible without requiring a separately retrieved derived fact. Existing relevance, provenance and context budgets still apply; this exemption grants no graph witness or source-expansion authority.
• Include the already-published OpenClaw 0.0.2 response-boundary fix: bound streamed JSON responses and escape recalled evidence in its wrapper. DSH remains at 0.0.1; these package versions are independent of the product version.
• Derive CLI version output from installed package metadata and require completed, current documentation reviews for version bumps and tagged candidate builds.
• Correct stale operational documentation about readiness, semantic work handoff, Profile storage, administrator user creation and typed memory removal. Deterministic L2 concurrency fixtures now include the required Insight title.

Upgrade

Apply all migrations with the operator migration command before starting upgraded APIs and workers. Migration 0036 backfills missing current Insight titles; migration 0037 adds/backfills accepted-input activity and the isolated administrator user-list read path. These metadata backfills do not require reindexing or model calls. Historical activity uses available accepted-version timestamps and is not an exact commit-time reconstruction. Existing originals and compatible completed indexes remain reusable. The product version is 0.0.2, while the existing contract_version: "0.0.1" and /v1 HTTP paths remain unchanged.

Use the product 0.0.2 image with the existing independent integration versions: OpenClaw 0.0.2, MCP 0.0.1 and DSH 0.0.1. Existing npm versions are reused, not published again. Keep operator secrets/configuration separate from public artifacts.

Verification and limitations

The isolated default suite accounts for all 2,395 collected test identities: 2,394 passed and one optional populated-scope read-only performance check was skipped. These are disjoint resumed runs after a local file-open stall, not one uninterrupted pytest invocation. Another 34 browser tests passed. OpenClaw, MCP and DSH suites passed 67, 28 and 11 tests respectively. Ruff, formatting, mypy, locked resolution, Python distribution builds and the 29-document release gate passed.

Both Linux ARM64 and AMD64 images report product 0.0.2 and match all 178 packaged source/configuration/migration files. ARM64 ran natively; AMD64 startup, authorized memory reads, administrator user listing and Inspector delivery passed under emulation. The native ARM64 image passed the real-provider L0/L1/L2, cited fast/normal recall, explicit Profile refresh and zero-call unchanged reuse, HTML/PDF/PNG original preservation and joint recall checks. Ten concurrent reads reached both API replicas. Authorized Inspector data and Insight titles loaded with required assets returning 200 and no JavaScript errors. User-list authorization/filter/validation and idempotent activity checks passed; downloaded portable-export members passed byte-length/SHA256 checks, title inspection and native export validation. The published MCP 0.0.1 archive ran in a fresh stdio process and returned cited recall and the stored Profile.

The first real-provider fast recall returned 503; repeated identical requests and the resumed candidate probe passed without reindexing. Its cause remains unverified, and the failed attempt and resume are retained. The probe used 19 generation requests: nine recorded requests consumed 22,215 generation tokens, including 6,707 reasoning tokens already included in output totals; ten preflight requests did not report usage. Embedding usage is excluded from these generation totals, and monetary cost is unknown. After publication, record the image digests, anonymous-download checksums and final source commit with the GitHub Release artifacts.

The independently versioned integration archives match their reviewed runtime bytes. Fresh full OpenClaw and DSH host capture loops were not repeated for this product release; their package suites and archive checks do not replace that boundary.

The tokenizer optimization already exists in the v0.0.1 source tag and is retained. Historical LME-500 and BEAM-10M results are not fresh measurements of v0.0.2; no accuracy, latency or token gain is claimed from this release. Physical multi-host failure, database high availability, backup restoration and adoption-scale user-list query cost remain unverified. The user-list page ceiling is not a performance guarantee.

Publication

Source commit: da89864ae1dbbb288157e86a7cdaf80b95abdf72 (annotated tag v0.0.2). Hosted documentation checks passed for both main and the tag.

Image (Linux AMD64 and ARM64):

ghcr.io/rocke2020/open-agent-memory@sha256:3412f1faafdaf6afad7026ded345fb17417942aafd0d643f2420bc3fb58ceaf1

Version tags 0.0.2 and v0.0.2, plus aliases 0.0, 0 and latest, resolve to this manifest. Anonymous pulls match the exact tested architecture images. Fresh processes from the published digest passed migrations, authorized reads, idempotent retain, cited joint recall, zero-call unchanged Profile refresh and Inspector asset checks on native ARM64 and emulated AMD64. The earlier v0.0.1 image is preserved.

Assets include the tagged source, Python wheel/source distribution, existing independently published integration archives, sanitized evidence and SHA256SUMS. No npm version was republished.

Public-read verification completed: all eight release assets were downloaded without credentials through GitHub’s public asset API and match their SHA256 checksums. The attached source archive matches all 481 files in the tagged tree.

Open Agent Memory v0.0.1

OAM v0.0.1 preserves cited interaction evidence, supports learned memory and Insights, provides explicit Profile refresh, and keeps document Evidence separate from learned memory. It includes the Memory Inspector, public OpenClaw and DSH npm packages, and PostgreSQL-backed API/worker deployment assets.

Source: c0f2d2e1f62decf3af188d7c5dff2a529134610e (v0.0.1). Indexing and recall design and acknowledgments ship with this release. Both plugin archives are the exact bytes already published as @rocke-dong/oam-openclaw@0.0.1 and @rocke-dong/oam-dsh@0.0.1.

The release source archive, plugin archives, sanitized evidence and SHA256SUMS are attached. Candidate evidence retains its original source and measurement identities. Historical evaluation results are development-snapshot evidence, separate from current native diagnostics; no campaign was rerun for publication.

The public container supports Linux AMD64 and ARM64:

docker pull ghcr.io/rocke2020/open-agent-memory:0.0.1

Immutable image: ghcr.io/rocke2020/open-agent-memory@sha256:5943b6c0f7a9c288933706396d841df7435b3415473e42f57e6fe50c2c9f383c. Tags v0.0.1, 0.0, 0 and latest resolve to the same digest at publication. The image includes the API, worker, operator commands and Memory Inspector; configure PostgreSQL/pgvector and model endpoints separately. Set OAM_IMAGE to this image in deploy/.env and follow the initialization/start commands in the Compose guide; building the candidate locally is unnecessary when using this published image.

Anonymous pulls passed on both architectures. Against this digest, the real HTTP path verified retained sources, learned memory, supported Insights, cited fast/normal recall, explicit Profile refresh with unchanged zero-call reuse, HTML/PDF/PNG Resources and joint recall. Original document downloads independently matched their accepted SHA-256 hashes. The authorized Inspector loaded real data and required assets without browser errors. Fresh native installations of both public npm packages completed capture, readiness, cited automatic recall in separate sessions and explicit native search. AMD64 startup and stored-data/asset reads passed under emulation.

OAMB main and dev at b03be82c41f53a4e58b6aaf2f51806ae9d441859 select the final OAM commit above. Its isolated native LME and BEAM diagnostics pass 1/1 each and reuse completed artifacts on resume; these diagnostics do not replace the retained full historical evaluations. The evidence ZIP distinguishes earlier candidate checks from current published-image checks.

Deployment acceptance is Mac-local simulation: ARM64 native and AMD64 emulated. Physical multi-host failure, database HA and remote CI are unverified. Original document bytes are preserved, while OCR accuracy is not guaranteed. Complete monetary cost remains unknown.

Public download acceptance completed at 2026-11-08T11:48:52.152626+00:00: unauthenticated repository, Release and annotated-tag reads return 200; all five Release assets download without credentials and match their reviewed bytes and SHA256SUMS. Both the attached source archive and GitHub-generated tag archive match all 448 tracked files at the final commit. The evidence ZIP preserves its earlier measurement timestamps, including the pre-visibility snapshot.