ai-memory-hub 1.0.1 beta
Released: 8 November 2026
ai-memory-hub 1.0.1 beta expands the hub from core memory storage and retrieval into a practical continuity layer for AI coding agents. The release adds safe conversation imports, durable handoffs, stronger retrieval quality, external OpenID Connect (OIDC) support, and an optional two-tool MCP surface. It also includes substantial performance, reliability, and supply-chain hardening.
Highlights
Import existing agent conversations
The hub can now import conversation history from Codex, Claude Code, GitHub Copilot CLI, Microsoft Copilot, OpenCode, Gemini CLI, Qwen Code, Pi and Oh My Pi, legacy OpenClaw, Hermes Agent, DeepSeek web shares, DeepSeek Harness, and Factory Droid. Imports are available through the CLI, the Connect UI, and the authenticated multipart HTTP endpoint.
Imported sessions have stable, project-scoped identities where the source provides a reliable session ID. Re-importing a growing session can append its new tail safely. Exact retries deduplicate, older snapshots are reported as stale, and diverged histories are reported as conflicts instead of being silently merged.
Use source discovery for a supported local agent:
uv run aim import latest codex --non-interactive --json
uv run aim import latest claude-code --non-interactive --json
Or select a parser and file explicitly:
uv run aim import codex-rollout-jsonl /path/to/rollout.jsonl --json
uv run aim import opencode-session-json /path/to/session.json --json
API clients can upload the original file without placing its contents in model context:
curl -fsS -X POST http://127.0.0.1:8000/memory/import \
-H "Authorization: Bearer $MEMORY_API_TOKEN" \
-F "parser=codex-rollout-jsonl" \
-F "file=@/path/to/rollout.jsonl;type=application/x-ndjson"
Uploads are bounded, require an explicit registered parser, and return indexed receipts without echoing transcript text, embeddings, content hashes, source paths, or provider exception details. See Import Conversations and Zero-token Import Hooks for source-specific formats and automated closed-file recipes.
Save and resume agent handoffs
Handoffs let one agent save unfinished work and another resume it with cited context. They use the normal validation, authorization, deduplication, audit, and storage paths. Updates are append-only continuations rather than rewrites of existing memory.
uv run aim handoff create conversation.json --json
uv run aim handoff search "release checks" --json
uv run aim handoff get <memory-id> --json
uv run aim handoff update <memory-id> continuation.json --json
MCP clients can use the create_handoff and resume_handoff prompts for the
same workflow. Optional source_client and source_session_id metadata help a
receiving client locate the correct work without weakening owner or project
permissions.
More precise search and answers
Search and memory_ask now share typed query planning for source, date, tag,
thread, handoff, intent, and temporal constraints. Retrieval combines dense,
lexical, fact, and graph evidence with reciprocal rank fusion and supports an
optional bounded semantic reranker with deterministic fallback.
Fact extraction now validates source spans and calibrates confidence. Answering arbitrates fact and direct-memory evidence, preserves correction precedence, and exposes conflict provenance. Capability-aware vector filtering pushes trusted owner and project constraints into providers that support it, with a bounded metadata-validated fallback for those that do not.
Existing commands continue to work:
uv run aim search "local-first tools" --top-k 5 --json
uv run aim ask "What did I store about local-first tools?" --top-k 5 --json
External OIDC resource servers
Deployments can use an external OIDC provider such as Keycloak. The hub uses discovery and cached JWKS data to validate asymmetric access tokens, qualifies identities by issuer, maps provider roles to memory scopes, applies configured allowlists, and publishes MCP authorization metadata.
api:
auth: oidc_resource_server
oidc:
issuer: https://id.example.com/realms/agents
audience: ai-memory-hub
Provider-specific setup, claim mapping, and Keycloak examples are in External OIDC and Keycloak.
Optional simple MCP profile
The default advanced MCP profile retains the full tool surface. Agents that
benefit from a smaller surface can opt into the simple preview, which exposes
only memory_save and memory_ask:
mcp:
tool_profile: simple
memory_save is an alias for the existing memory_insert handler. The profile
does not create a second write path or remove the underlying HTTP, retrieval,
fact, or profile capabilities.
Performance and reliability
- Exact duplicate ingestion now skips derived-record, embedding, metadata, and vector work when the stored index is complete.
- Trusted appends reuse verified conversation identity, summarize only the combined result, and limit graph extraction to newly appended messages.
- Remote embedding batches are bounded by both input count and UTF-8 byte size.
- Retrieval avoids repeated source hydration, redundant search-text assembly, trivial reranker calls, and exact adjacent context overlap.
- Deterministic quality, scale, concurrency, latency, and work-amplification reports make provider and model changes easier to compare.
- Import reconciliation and receipts make scheduled imports safely repeatable.
Security and dependency updates
This release requires verified OIDC email claims before email-domain allowlists are applied, bounds complex Markdown and imported JSON contracts, keeps docs preparation dependency-light, and includes dependency and container updates for reported security advisories. Secrets, transcripts, embeddings, and full user queries remain excluded from operational logs.
Upgrade from 1.0.0 or the 1.0.0 beta releases
- Back up the configured metadata and vector stores.
- Check out the release and refresh the locked environment.
- Review authentication and storage configuration before restarting.
- Start the hub and verify readiness, then run one insert/search or import/ask smoke test for the configured project.
git fetch --tags
git checkout v1.0.1-beta
uv sync --dev --locked
uv run aim serve --host 127.0.0.1 --port 8000
The project requires Python 3.14. The default local setup remains SQLite plus
LanceDB. Existing advanced MCP configuration remains compatible because the
new simple profile is opt-in.
Release artifacts
- Source tag:
v1.0.1-beta - Documentation: https://artemon-line.github.io/ai-memory-hub/
- Container:
ghcr.io/artemon-line/ai-memory-hub:1.0.1-beta - Full change list: Changelog
The container publishing workflow adds the immutable image digest to the GitHub release after the image has been built, scanned, pushed, and smoke-tested.
This is a prerelease. It does not update the container latest tag and should
be promoted to a stable release only after the release-candidate checks and
clean-environment smoke tests pass.