Agent Memory Systems

Agent memory gives LLM agents state that outlives a single context window: user preferences, facts, past episodes. It overlaps with RAG (retrieval-augmented-generation-overview) but the store is written by the agent’s own interactions, so it needs extraction, updating and forgetting, not just search. See also context-engineering.

Main open-source options (per GitHub READMEs, 2026-10-02)

• Mem0 (Apache 2.0, about 66.6k stars; latest release ts-v3.3.1, 2026-09-25): memory layer with user, session and agent levels, self-hosted server or cloud platform, CLI. Retrieval combines semantic, BM25 and entity matching. Its README cites a “new memory algorithm (April 2026)” scoring 92.5 on LoCoMo, 94.4 on LongMemEval and 64.1 on BEAM (1M tokens); these are vendor-reported and not independently verified. Paper: arXiv 2504.19413.

• Letta (Apache 2.0, about 25k stars; formerly MemGPT; last tagged release 0.16.8 on 2026-05-14, last push 2026-09-10): stateful agents with editable memory blocks. The README says active development has moved to the letta-ai/letta-code repository (agent harness, app server, channels such as Slack/Telegram/Discord); the older server code is preserved on an archive branch. See letta.

• Zep / Graphiti: Graphiti (Apache 2.0, about 31.5k stars; v0.30.2 on 2026-09-08) builds temporal context graphs where facts carry validity windows and provenance back to source episodes, with hybrid semantic, keyword and graph retrieval. Zep is the managed platform built on it. Backends: Neo4j 5.26+, FalkorDB, Amazon Neptune (Kuzu marked deprecated). Paper: arXiv 2501.13956. See Neo4j.

• LangMem (MIT, LangChain, about 1.7k stars): core memory API plus LangGraph storage integration; “hot path” memory tools the agent calls mid-conversation and a background manager that extracts and consolidates memories. No tagged GitHub release; last push 2026-10-02.

• Cognee (Apache 2.0, about 31k stars; v1.6.2 on PyPI 2026-09-29): knowledge-graph memory with remember / recall / improve / forget operations. See cognee.

• Pixeltable is, in the author’s view, adjacent (a multimodal data backend usable for agent memory) rather than a dedicated memory layer. See pixeltable.

Managed plans (Mem0, mem0.ai/pricing, 2026-10-02)

Hobby free (10k add / 1k retrieval requests a month); Starter; Pro (graph memory, analytics, “Dream” memory consolidation); Enterprise custom (SSO, on-prem, audit logs). Zep and Letta Cloud plans not checked. Current prices: see the vendor pricing page https://mem0.ai/pricing

Design axes

Explicit editable blocks (Letta) vs extracted fact stores (Mem0) vs temporal graphs (Zep/Graphiti); write on the hot path vs background consolidation; scoping per user, session and agent. Related: hebbian-vault (organisational memory), graphrag.

Caveats

LoCoMo and LongMemEval numbers are self-reported by the vendors (methodologies may differ; not verified); run your own evaluation (rag-evaluation).

Open items

• Mem0, Letta, Graphiti and LangMem benchmark and design claims come from vendor READMEs only; not independently verified.
• Mem0 and Zep papers (arXiv 2504.19413, 2501.13956) not opened.
• Zep managed pricing and current state not checked.

Related

• Beads - git-backed task/memory tracker for coding agents

Sources

• https://github.com/mem0ai/mem0 and GitHub API repo/release metadata for mem0, letta, graphiti, langmem (accessed 2026-10-05)
• https://github.com/letta-ai/letta (accessed 2026-10-02)
• https://github.com/getzep/graphiti (accessed 2026-10-02)
• https://github.com/langchain-ai/langmem (accessed 2026-10-02)
• https://mem0.ai/pricing and https://github.com/topoteretes/cognee (accessed 2026-10-02)
• Mem0 paper https://arxiv.org/abs/2504.19413 and Zep paper https://arxiv.org/abs/2501.13956 (ids as cited by the READMEs; papers not opened)