Skip to content

v2.0.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 02 Sep 16:07
· 8 commits to main since this release
v2.0.0
b77b653

ai-memory 2.0

The wiki is now a portable open format, retrieval is hybrid by default, and memory quality is measured — LongMemEval-S hit@5 went from 0.617 → 0.823 in this release.

Highlights

  • Local embeddings, on by default — in-process sentence embeddings (pure Rust, no API key, no external server). The model (~87 MB) downloads in the background on first start and existing pages are backfilled automatically; hosts that can't fetch it keep working FTS-only. Opt out with embedding_provider = "none". → docs/local-embeddings.md
  • Open Knowledge Format v0.2 on disk — every page is a conformant OKF document; ai-memory export-okf streams a validated bundle you can import anywhere. → docs/okf.md
  • Time-travel queries — memory_query accepts as_of to answer "what did we know about X then", including knowledge superseded since. → docs/temporal.md
  • Typed relation edges — causes / fixes / contradicts in page frontmatter; declared contradictions surface as lint findings with zero LLM cost. → docs/typed-edges.md
  • Cross-session experience pass (opt-in) — every N sessions, recent trajectories are reviewed side by side for knowledge visible only across sessions, staged through the same gated auto-improve path. → docs/experience.md
  • Retrieval benchmark harness — ai-memory-eval retrieval runs LongMemEval-S against a real server end to end; baselines published. → docs/benchmarks/
  • Search quality fixes — natural-language queries no longer crash FTS5 or surface stopword-only matches; status now reports embedding coverage, wiki format, typed edges, and write-queue backpressure truthfully.
  • Offline durability for TypeScript integrations — OpenCode, OMP, Pi, and OpenClaw now spool failed hook deliveries and drain them when the server is reachable again, exactly like the shell hooks (#580).
  • Also: single-instance serve lock (#563), refusal of misfiled unscoped writes (#564), macOS launchd agent (#569), llm_reasoning_effort across providers (#557), atomic capture-mode writes, and a laptop-offline day's worth of hardening. Full list in CHANGELOG.md.

⚠️ Migration — read before upgrading

  • First 2.0 start migrates the wiki to OKF in place. It is gated on a verified full backup of the data dir, archived to your home directory (or AI_MEMORY_BACKUP_DIR); if the backup cannot be written and verified, the server refuses to start. Allow disk space and a moment proportional to your store.
  • One-way door for shared stores. Once migrated, pre-2.0 binaries refuse the data dir. If several machines share one store, upgrade all of them; don't downgrade.
  • Default embeddings change behavior. First start downloads the model in the background and backfills existing pages; hybrid search enables on the next restart. Air-gapped installs stay FTS-only (or drop the pinned model files in manually).
  • Container/wrapper installs have specific path notes — see docs/MIGRATION-2.0.md for the full checklist.

Full changelog: v1.39.0...v2.0.0

Checksums (SHA256)

59c604813b74d7256ccc2e09007d5cb025106f360ff35a2112349b6df7e03e4c  ai-memory-hooks.tar.gz
448f6cfad616e13692598dd49709ce9dec10eed0f702818ccca4a0c54a79f8c7  ai-memory-install-hooks
d889eb21d242d7409ce06729c3f8aa506380d19e0ad02284ef9aef50a2cc5b4d  ai-memory-linux-aarch64.tar.gz
b872b162104dadbe12e7b1498c0e6e439709d7fac25859e183a9c19e7bf6ced9  ai-memory-linux-x86_64.tar.gz
3433db418fd1a3573dfda0ccedf1f7f8a40820d5af3b96912d5944bfd153d194  ai-memory-macos-aarch64.tar.gz
9d0d018fce6d53a744f64cc4ba27fd240c5f0c5a3ec1727eeb72fbe729091031  ai-memory-macos-x86_64.tar.gz
abfccc8cbb53108201614921196e069fc16f0b07bb55aa6c22bcafedb13295b3  ai-memory-windows-x86_64.zip
6a91c44ffa2e85d3b5a6ce26a4ccff0286b91d08d276d2bb4d47c346b036ee8b  ai-memory-wrapper.cmd
f4912fadae3f7aa0fa2f1612b40c8a915462330c217a6e7780cdd7be5ce5b068  ai-memory-wrapper.ps1
4b2e5736195f0ac4cd38adf62f25b294922ebf81f5f4802a07803e1abbf9f72d  ai-memory-wrapper

Install

# Arch Linux (AUR)
yay -S ai-memory-bin      # prebuilt Linux x86_64/aarch64 binary
yay -S ai-memory          # builds from source

# Docker
docker pull akitaonrails/ai-memory:2.0.0

# macOS aarch64/x86_64 (no toolchain): download the matching ai-memory-macos-<arch>.tar.gz below,
# extract it, and follow the bundled docs/macos.md.

# Windows x86_64 (no toolchain): download ai-memory-windows-x86_64.zip below,
# extract it, and follow the bundled docs/windows.md.

# From source
cargo install --git https://github.com/akitaonrails/ai-memory --tag v2.0.0
Assets 22
Loading