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

Memory

Author
Tom Rochette
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: in progress / Readability: 3 / 4 minutes read (~763 words)
agents 48 memory 28
Table of Contents

Persistent memory for agents: the file conventions, the portable format, the capture plugins, and the hosted and self-hostable services, from graph pipelines to temporal stores.

  • agentmemory - the Apache-2.0 keyless session-memory server shared by 20-plus coding agents: BM25 recall first, 54 MCP tools, on a pinned iii engine.
  • Basic Memory - the AGPL-3.0 MCP server that grows a knowledge graph from agent conversations in Obsidian-compatible markdown, with per-seat Team and Business cloud tiers.
  • Cabinet - the MIT self-hosted knowledge base where every artifact is a markdown file on disk and an onboarded agent team works it on a schedule, quiet since August.
  • claude-mem - the 100.2k-star plugin that captures coding-agent sessions, compresses them with your tokens, and reinjects the context.
  • Cognee - the Apache-2.0 graph-memory pipeline with the whole engine self-hostable and a flat per-token cloud.
  • deja - the MIT one-binary transcript index: recall of what 41 coding agents already wrote to disk, no capture step, no LLM, self-run benchmarks committed as artifacts.
  • Emulo - the MIT session miner that compiles the messages you typed into a you.md working profile with dated verbatim receipts, carrying a published 0-for-9 placebo result.
  • Engrim - the local-first SQLite episodic memory engine for multiple AI CLIs on one machine, provenance-first.
  • EverOS - EverMind’s Apache-2.0 local-first runtime where markdown files stay the source of truth with SQLite and LanceDB indexes beside them, separate user and agent tracks, and an unpublished-price managed cloud.
  • File-based agent memory - the CLAUDE.md and AGENTS.md conventions, memory as plain markdown files.
  • GBrain - Garry Tan’s MIT personal brain daemon: provenance-tagged markdown pages with corrections and withdrawal, MCP-served to any agent, with published eval receipts.
  • Hindsight - Vectorize’s MIT memory engine (facts, observations, knowledge pages) behind retain, recall, and reflect, 49.0k stars, with per-run benchmark artifacts and a pay-as-you-go cloud.
  • Letta - the MemGPT creators’ memory-first platform, agent plus cloud tier.
  • LongMemory - CaviraOSS’s Apache-2.0 temporal memory engine: immutable, provenance-carrying memories in SQLite with point-in-time recall and governed scopes, self-hosted with MCP and a dashboard.
  • mem0 - the hosted and self-hostable memory layer across vector, graph, and key-value backends.
  • Memanto - Moorcheh’s MIT memory agent that manages a fleet’s memory on a schedule: 13 typed categories, contradiction review, policy expiry, and OKF markdown export.
  • MemMachine - MemVerge’s Apache-2.0 memory layer pairing graph-stored episodic memory with SQL-stored profile memory, self-hosted or on a free cloud account.
  • Memoryfields - Cal Paterson’s portable memory-as-file-format spec, markdown pages plus a deletable vector index, the files-over-pipelines argument in RFC form.
  • MemOS - MemTensor’s Apache-2.0 memory OS with official OpenClaw, Hermes, and DeepSeek Harness plugins, a Neo4j-plus-Qdrant self-host, and a hosted API listing $19 and $286 tiers as launch-free.
  • memU - the Apache-2.0 memory wiki that turns agent history into markdown skills shared across ChatGPT, Claude Code, Cursor, and OpenClaw, hosted free with a single-device self-host.
  • Memvid - Memvid’s Apache-2.0 single-file memory engine: documents, embeddings, and indexes packed into one portable .mv2 file, the QR-in-video v1 now deprecated.
  • MetaClaw - the MIT meta-learning layer that evolves OpenClaw-family agents’ skills and memory from conversations, dormant since June.
  • OpenViking - ByteDance’s AGPL context database that unifies knowledge, memory, and skills as a viking:// filesystem with layered summaries, 39.8k stars.
  • PLUR - the Apache-2.0 local-first engine giving every wired harness one shared store of plain-text YAML engrams, with auto-injection adapters for eight coding agents.
  • Screenpipe - Negentropy Labs’ source-available capture engine (YC S26): a local screen-and-audio timeline agents query over MCP, with paid commercial licensing.
  • Supermemory - the benchmark-forward memory API (facts, profiles, and RAG in one engine) with an MIT one-binary local mode and a $19-$399/mo hosted ladder.
  • TencentDB Agent Memory - Tencent’s MIT plugin giving OpenClaw and Hermes reversible compression into node_id drill-down Mermaid canvases over a local persona pyramid.
  • Zep - temporal knowledge graphs where contradictions invalidate old facts.

Its members are compared on shared rows in the Memory Feature Matrix.

Changes
#

  • 2026-08-24 - Added File-based agent memory.
  • 2026-08-24 - Added Letta.
  • 2026-08-24 - Added mem0.
  • 2026-08-24 - Added Zep.
  • 2026-08-26 - Added Cognee.
  • 2026-08-30 - Added claude-mem.
  • 2026-09-04 - Added Memoryfields.
  • 2026-09-10 - Added Engrim.
  • 2026-10-02 - Added Cabinet.
  • 2026-10-02 - Added MetaClaw.
  • 2026-10-04 - Added Supermemory.
  • 2026-10-06 - Added Basic Memory.
  • 2026-10-06 - Added MemOS.
  • 2026-10-06 - Added memU.
  • 2026-10-07 - Added GBrain.
  • 2026-10-07 - Added Hindsight.
  • 2026-10-07 - Added OpenViking.
  • 2026-10-08 - Added LongMemory.
  • 2026-10-08 - Added Memvid.
  • 2026-10-08 - Added Screenpipe.
  • 2026-10-09 - Added agentmemory.
  • 2026-10-09 - Added EverOS.
  • 2026-10-10 - Added deja.
  • 2026-10-11 - Added Memanto.
  • 2026-10-11 - Added MemMachine.
  • 2026-10-11 - Added TencentDB Agent Memory.
  • 2026-10-11 - Added Emulo.
  • 2026-10-14 - Added PLUR.