↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Memory/

LongMemory

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 6 minutes read (~1086 words)
agent-curated 587 fully-ai-generated 690 llm=glm-5.3-flash 597 mcp 30 memory 34 research-note 557 self-hosted 25 sqlite 4 temporal 1
Table of Contents

LongMemory is an Apache-2.0 local-first memory engine: a TypeScript library, CLI, HTTP server, and MCP server over SQLite that keeps immutable, provenance-carrying memories with separate recorded and valid time, governed scopes, and four recall modes.

Its bet is that memory needs the database properties agents never get (immutability, temporal truth, governance, explanation), and it is the first local option in this category to build all four into the open engine.

What it is
#

A project by CaviraOSS, renamed from OpenMemory (the old openmemory-js and openmemory-py packages remain as forwarding bridges; the official distribution is longmemory on npm, at 1.3.3). The Hydrograph engine keeps immutable nodes, typed executable edges, worlds, entities, and traces, persisted to SQLite. Recall comes in four modes: strict (temporal, contradiction, confidence, and grounding gates), historical (point-in-time queries with a supplied valid time), associative (semantic, lexical, entity, and graph signals), and world-grounded (requires current external evidence). Governance is structural: tenant, organization, project, user, team, role, agent, task, and framework scopes, with MCP tool arguments unable to override server-bound identity. Surfaces include the npm library, a CLI, an HTTP API on port 7331, an MCP server with 13 governed tools, a Next.js dashboard, a VS Code extension, a session porter that imports Claude Code, Codex, OpenCode, Gemini CLI, Copilot, and Cline logs, an n8n community node, and one-click templates for Docker, Heroku, Railway, and Render.

Status
#

Young but shipping, with a strong independent launch record for its size. 4,527 stars as of 2026-10-11 (GitHub API), with the repository pushed 2026-10-18, created 2025-10-19, as of 2026-10-29 (GitHub API).

Star History Chart

The Show HN thread (December 2025, as OpenMemory) drew 48 points and 16 comments, and the contributor count stands at 51 by GitHub API pagination as of 2026-10-29 (the README carries a contributors table again, 37 avatars). Its benchmark harness publishes manifests and fails closed on incomplete datasets or embedding fallback, covering LongMemEval, LoCoMo, and BEAM, though the published results are still the project’s own runs. The README has grown well past the released npm train: it now documents multilingual recall (script detection, code switching, transliteration), an answer_from_evidence adapter making at most one model call over validated citations, governed assets (Chat Memory, Skills, LLM-Wiki, CodeGraph), a project-memory CLI with skill matching and asset loadouts, an Agent Plugins 1.0 bundle for OpenClaw, plugins and packs for Claude Code, Codex and ChatGPT desktop, Gemini CLI, Cline, Continue, LibreChat, Dify, and Flowise, and deployment targets for DigitalOcean, Vercel, and Windows, while npm stays frozen at 1.3.3 since August 31, so the newest surface ships docs-side, not release-side. Two verification gaps carried into this run: the PyPI distribution the README instructs you to install (longmemory-sdk) returned 404 again on 2026-10-29 (as on every check since 2026-10-08), and the repository has been quiet since the October 18 push.

Strengths
#

  • Temporal truth in the open engine: recorded time and valid time are separate, with supersession and point-in-time recall, territory every other local option here leaves to Zep’s paid tier.
  • Immutable storage with hashes and provenance, so recall and decay cannot rewrite history.
  • Deterministic governance with server-bound identity, the strongest multi-scope story among the self-hostable members.
  • The benchmark harness’s fail-closed manifests refuse to publish incomplete runs, an unusually careful design for a self-run scoreboard.

Cautions
#

  • Young and quiet: nothing pushed since 2026-10-18, an npm train frozen at 1.3.3 since August 31, and a core team small enough that bus factor is a fair question; the README’s expanding feature list is ahead of any released artifact.
  • The Python SDK distribution 404’d on both the JSON API and the simple index this run (again, 2026-10-29), so the multi-language story is uneven in practice.
  • The rich engine wants operations: embeddings need provider keys, and the server, dashboard, and extension paths add surface the in-process library hides.
  • All published benchmark results are self-run; no independent replication exists yet.

Pricing
#

Pricing does not apply: Apache-2.0 (the n8n node MIT by that registry’s rules), self-hosted, no hosted tier, as of 2026-10-08.

Compared to
#

  • Zep: the hosted temporal-graph service; LongMemory is the first open engine here with point-in-time truth, though unproven at Zep’s production scale.
  • Cognee: the other self-hostable graph option; Cognee builds graphs from documents in pipelines, LongMemory keeps a governed temporal store agents read and write.
  • Engrim: the other local cross-agent store; Engrim is a thin curated SQLite scratchpad, LongMemory is a full engine with governance and temporal queries.

Bottom line
#

Recommended for teams that need temporal, governed memory they can self-host and will accept a young dependency. Not for anyone needing a hosted tier, a proven-at-scale engine, or a Python SDK that installs today.

Changes
#

  • 2026-10-08 - Created from the 2026-10-08 awesome-list entrant scan, with six fetched sources; the unresolvable PyPI SDK distribution recorded as the run’s verification gap.
  • 2026-10-18 - Rotation re-verification: the push drought ended with a community project-ingest embeddings fix merged 2026-10-17, so the quiet caution was revised; the contributor count moved to 45 by GitHub API after the README dropped its listing; the PyPI SDK 404 re-confirmed through 2026-10-18.
  • 2026-10-25 - Corrected the star-history chart fallback markup (the img fallback carried the dark-theme parameter; it now points at the default chart).
  • 2026-10-26 - Corrected the star-history chart markup: the light-theme source carried the dark-theme parameter; it now points at the default chart.
  • 2026-10-29 - Rotation re-verification: every cited URL re-fetched live (repository API and page, README, license, npm, the PyPI simple index, the Show HN thread), no facts reversed; the contributor count moved to 51 by API (the README carries a contributors table again), the README’s documented surface grew (multilingual recall, the answer_from_evidence adapter, governed assets, new integrations and deployment targets) while the npm train stays frozen at 1.3.3, the PyPI SDK 404 re-confirmed 2026-10-29, and the quiet record now reads as no push since 2026-10-18.

See also
#

  • Memory Feature Matrix - the category comparison this note joins
  • Zep - the hosted temporal service this engine’s temporal model answers
  • Cognee - the pipeline-style self-hosted graph alternative
  • Engrim - the thin local store at the opposite end of the complexity spectrum

References
#