AgentMemory 0.9.31 improves memory retrieval, reduces repeated model work, and fixes capture, recovery and import failures. It also adds native Pi, Oh My Pi and DSH integrations and upgrades the runtime to iii 0.25.0 with Worker Compose.
Retrieval and performance
- Better keyword ranking: fixing prefix-score inflation raised strict recall-all@5 from 65.74% to 84.68%, a gain of 18.94 percentage points, in the documented BM25 development evaluation. Complete-evidence retrieval increased from 309 to 398 of 470 answerable questions, with 92 recovered cases and 3 regressions. These are component retrieval results on inspected development data, not held-out answer accuracy or a release-wide speed benchmark. Methodology and receipts, #1514.
- Less repeated summary work: summaries process new observations and merge them into the existing summary. An unchanged session makes zero additional summarization calls and skips repeat reflection, graph extraction and skill processing. #1507.
- Bounded graph work: search shares indexed graph reads and avoids repeated full-graph enumeration. Optional graph enrichment has a 500 ms budget so ready keyword and vector results can return if graph storage stalls. This is a graph-stage budget, not an overall search-latency guarantee. #1541, #1559.
Fixes
- Session summaries cover later turns instead of stopping at the first Stop event. Idle sessions finalize, and abandoned sessions are summarized before abandonment. Fixes #745.
- Saved memories retain project and agent scope and remain recoverable when token budgets shorten or omit content. Recall can target memories, observations or both. Fixes #819 and #1232.
- Failed compression preserves searchable captures and supports scoped inspection and retry. Vector recovery retains unfinished work across checkpoints, provider failures and restarts. #1546, #1562.
- Replace imports validate data before deleting stored records. JSONL imports preserve session timestamps and avoid repeated duplicate observations; the first re-import of rows from older versions can still add one final copy. #1576, #1526.
- Hooks capture final assistant replies and preserve distinct subagent, task and notification events. #1510, #1513.
- CLI status shows the actual embedding provider, and doctor uses the daemon's LLM-provider detection. Fixes #1488 and #1572.
- MCP keeps routine logs off stdout and identifies writes that only reached local fallback storage. The viewer loads actions beyond the first page, and deleting memories cleans up their facets. #1522, #1563, #1524, #1591.
Integrations and documentation
Bundled Pi, Oh My Pi and DSH integrations provide native capture and memory workflows. Cursor resolves workspace ownership more reliably, and the Devin plugin captures lifecycle events correctly. Shared agent skills, setup documentation and README translations are updated. The bundled Pi connector works independently of the separate Pi package's npm publication.
Upgrading
Stop AgentMemory with the existing CLI before upgrading, then restart 0.9.31 using the same instance, port and data directory. The runtime now uses iii 0.25.0 with Worker Compose. Viewer updates use authenticated SSE on the REST port; port 3112 is unused. MCP narrative recall returns structured JSON; read its text field.
Re-run agentmemory connect pi --with-mcp or agentmemory connect omp --with-mcp to refresh bundled integrations. Existing DSH hook-bridge installations migrate with agentmemory connect dsh --with-hooks --force. On Windows, WSL2 is recommended; native Windows remains unverified and requires a manually installed pinned engine.
Upgrade guide | 0.9.31 changelog | All changes since 0.9.30
Thanks to our contributors
Thank you @guibeira, @JackieJK, @andrew05060414, @hejiawow, @dmazhukov, @Rex57, @SomSamantray, @yxtay, @akushonkamen, @DimitriChrysafis and @ErikBPF for your contributions to this release, and everyone who reported issues and tested fixes.