Changelog
ai-memory-hub uses this changelog as the committed release notes source. GitHub releases may additionally use generated notes, but the release summary here should describe the supported public behavior for each tagged release.
Unreleased
1.0.1-beta - 2026-11-08
See the 1.0.1 beta release notes for highlights, upgrade guidance, and copy-paste examples.
- Added
mcp.tool_profile:advancedremains the default full MCP tool surface, while the opt-insimplepreview registers onlymemory_saveandmemory_ask. The save name aliases the existing insert handler without changing HTTP endpoints or internal ingestion and retrieval capabilities. - Verified that DeepSeek web exposes per-conversation public shares and a
separate account-wide data ZIP, not a native per-conversation JSON download;
documented the observed contracts and added privacy-safe diagnostics that
reject account exports passed to
deepseek-share-json. - Raised the optional OAuth dependency floor to
joserfc>=1.7.6and locked1.7.7to address GHSA-9pg5-8c7x-5hfx. - Reconciled historical Markdown plans against shipped code and tests, moved remaining work into GitHub epics, and simplified public documentation around current features, evidence-based project status, and issue-backed planning.
- Standardized benchmark latency distributions and separated cold first-call measurements from warm p50/p95/p99 samples while preserving existing report fields.
- Added a shared request-work ledger to deterministic efficiency reports for comparable storage, embedding, candidate, reranker, audit, and context-work totals and per-request amplification.
- Added a scheduled and manually dispatched efficiency report suite that uploads machine-readable benchmark manifests and before/after reports without adding load profiles to normal push or pull-request CI.
- Added credential-free local concurrency profiles for insert, search, and ask, reporting completion, throughput, and latency percentiles at configurable worker levels against versioned synthetic corpora.
- Added deterministic small, medium, and large synthetic efficiency corpora with broad and selective filter cases, streamable JSONL output, and fingerprinted machine-readable manifests.
- Added a distinct Connect UI matrix for implemented and planned zero-token import hooks, with pinned versions, parser metadata, and safe documentation links.
- Added a bounded zero-token DeepSeek Harness ZIP import command with safe extraction, explicit descendant selection, and formats 0 through 4 support.
- Added an opt-in zero-token Hermes Agent 0.22.5 export adapter with private temporary files, optional redaction, ended-session batches, and cleanup.
- Added an opt-in OpenClaw compatibility adapter for closed legacy version 3 JSONL trees, with explicit refusal of current SQLite and sidecar files.
- Added an opt-in, explicit zero-token import adapter for closed Pi 0.73.0 version 3 sessions without newest-file guessing or background watchers.
- Added a shared zero-token conversation uploader with closed-file snapshots, bounded safe receipts, and opt-in Codex, Claude Code, OpenCode, and Gemini CLI import recipes.
- Bounded remote embedding batches by configured input count and UTF-8 byte size, preserving deterministic chunk mapping and all-or-nothing vector writes.
- Reused the already verified, project-scoped aggregate conversation identity for complete-history trusted appends, while retaining safe recomputation for sibling continuations whose payload omits part of the stored history.
- Limited trusted-append graph extraction to newly appended messages while preserving absolute message-index provenance and conflict supersession.
- Removed exact repeated token-window overlap from budgeted ask and handoff context when adjacent chunks are proven to come from the same source message.
- Skipped semantic reranker projection and invocation when zero or one candidate can be reranked, with an explicit diagnostic mode and deterministic call-amplification coverage.
- Adapted unfiltered lexical and fact candidate limits to the requested
top_kand configured candidate multiplier, while preserving configured maxima for filtered searches to protect post-filter recall. - Deferred generated-summary work until after trusted-append detection, so an append writes summaries once for the combined conversation instead of first writing transient summaries for the incoming transcript.
- Reused assembled source-conversation search text within fact ranking and question narrowing requests, with a deterministic work-amplification benchmark and no cross-request payload cache.
- Replaced repeated fact-source conversation reads during filtering, ranking, projection, candidate construction, and provenance assembly with bounded, deduplicated metadata batch hydration and a machine-readable work benchmark.
- Short-circuited fully indexed exact-duplicate ingestion before derived-record, embedding, metadata, and vector work, while retaining incomplete-index repair and adding a machine-readable ingestion work ledger.
- Reused bounded, file-identity-aware compiled conversation validators across ingestion requests, with explicit invalidation and isolated multi-schema runtime behavior.
- Improved performance benchmark fidelity with exception-free in-memory audit accounting, first-call and warm latency distributions, p50/p95/p99 reporting, and per-request candidate/context amplification metrics.
- Added a versioned synthetic retrieval and extraction quality corpus with a credential-free CI gate, comparable provider/model reports, auditable thresholds, ranking/filter/fact/temporal metrics, latency percentiles, and reranker fallback coverage.
- Added fact/direct-memory evidence arbitration for
memory_ask, including assistant-claim corroboration, cross-layer conflict provenance, corrected-fact precedence, and a high-confidence fact fast path. - Added source-span validation and calibrated confidence for extracted facts, including assertion/temporal classification, bounded external extractor output, assistant corroboration semantics, and legacy fact compatibility.
- Added provider-scale-independent reciprocal rank fusion for dense, lexical, fact, and graph candidates, exact-match prioritization, and an optional bounded redaction-safe semantic reranker with deterministic fallback.
- Added shared typed query planning for search and ask, including deterministic source, date, tag, thread, handoff, intent, and temporal parsing; safe optional model-assisted planning; explicit-filter merge rules; and opt-in diagnostics.
- Added capability-aware filtered vector search with trusted owner/project pushdown, bounded metadata-validated fallback for unsupported constraints, filter diagnostics, and a documented persistent-index reindex path.
- Added safe
aim import latest <source>discovery for Codex, Claude Code, GitHub Copilot CLI, Pi, and Oh My Pi, including documented environment overrides, active-file and ambiguity refusal, interactive preview, non-interactive JSON output, provider-matrix tests, and generated report evidence. - Added a generic external OIDC resource-server mode with discovery/JWKS caching, asymmetric access-token validation, issuer-qualified identities, Keycloak role-to-scope mapping, allowlists, MCP authorization metadata, and cached provider availability in readiness and Connect UI diagnostics. Valid external identities now synchronize configurable username, email, and display-name claims into stable local user profiles.
- Added authenticated multipart
POST /memory/importsupport for every registered conversation parser, with explicit parser selection, bounded uploads, safe indexed receipts, shared insertion policy enforcement, and a declarative versioned schema-override contract for Hermes exports. - Added a schema-backed Hermes Agent JSONL importer for single-session and multi-session exports, preserving safe provenance and live conversational text while excluding system, tool, reasoning, rewound, synthetic summary, billing, usage, and identity data.
- Added a versioned Factory Droid Exec capture importer for prompt/result and bidirectional JSON-RPC recordings, preserving completed visible turns while excluding reasoning, tools, errors, usage, and control frames.
- Added a schema-backed DeepSeek Harness canonical session JSONL importer for extracted root and descendant logs, preserving safe session provenance while excluding internal context, reasoning, tools, compaction, usage, and attachments.
- Added a schema-backed GitHub Copilot CLI
events.jsonlimporter that keeps top-level user and assistant text while excluding synthetic prompts, sub-agent traffic, reasoning, tools, compaction, usage, and telemetry. - Added a native Qwen Code session importer for
/export jsonand/export jsonlwith schema-backed message validation, UUID deduplication, and omission of system, goal-state, tool, usage, and unknown records. - Added a native Pi-family session JSONL importer for Pi, Oh My Pi, and archived/legacy OpenClaw transcripts with active-branch reconstruction and schema-backed tree validation.
- Added a native Gemini CLI session JSON importer for
/export-session,/chat shareJSON, and compatible legacy saved-chat objects while omitting injected context, thoughts, tools, token usage, and diagnostics. - Added a native Claude Code session JSONL importer that keeps visible user and assistant text while omitting reasoning, tools, system records, usage, compaction summaries, and file snapshots.
- Added a native Codex CLI/app rollout JSONL importer, verified against Codex CLI 0.152.0, that keeps canonical user and assistant text while omitting reasoning, tools, injected startup context, and runtime-only events.
- Added a native OpenCode session-export JSON importer, including CLI support and a live OpenCode/Ollama export, hub insert, and cited query round-trip smoke test.
- Updated the container PCRE package and locked urllib3 release to their fixed versions after the supply-chain scan detected new high-severity advisories.
1.0.0-beta
- Added first-release governance, support, and release automation scaffolding.
- Fixed detailed
memory_retrieveindex manifests so successfully embedded chunks reportindexedinstead of stalepending_index. - Fixed
memory_askrouting so active latest fact projections are preferred before direct chunk fallback, including corrected temporal facts. - Fixed
memory_profile_getsummary text so duplicate active fact rows are collapsed into canonical human-readable lines without mutating the timeline. - Patched the Bruno test toolchain against
js-yamlCPU-exhaustion andcsv-parseprototype-replacement vulnerabilities.
0.1.0
Initial public release candidate scope:
- HTTP memory API and streamable HTTP MCP tools.
- SQLite/LanceDB local default storage.
- Postgres/PGVector runtime option.
- Deterministic ingestion, search, retrieve, ask, facts, and summaries.
- CLI and container runtime.
- Bearer-token auth and project workspace boundaries.
- Local-first, bring-your-own embedding model/storage posture.
Install and run from source:
uv sync
uv run aim serve --host 127.0.0.1 --port 8000
curl -fsS http://127.0.0.1:8000/ready
Run the release image after Docker Hub publishing is configured:
docker run --rm -p 8000:8000 docker.io/<namespace>/ai-memory-hub:v0.1.0
curl -fsS http://127.0.0.1:8000/ready
Release notes must be updated with the final image digest after the publish
workflow completes. The documentation site is published from the checked-in
MkDocs configuration through the pages workflow.
Known limitations:
- No hosted memory service is included.
- Production-quality retrieval requires a bring-your-own embedding model.
- Browser extensions are planned as separate adapters and are not part of this release.
- UI dashboards and SDKs are future work.