Agent memory layers, ranked
Tools that give an agent long-term memory: they store conversations, project context and decisions so the next session, or a different agent, can recall them. Some are plugins for Claude Code or OpenCode, others are standalone services or MCP servers; some use vector search, others knowledge graphs. To be listed, cross-session memory must be the project's main job (assistants or frameworks that merely include memory are left out), with at least 2,000 GitHub stars and a commit in the last 90 days. Memory often holds code and personal data, so check where it is stored and whether it is sent to a third party.
As of 2026-11-09, the top-ranked agent memory layer by FARS source review is Cognee Memory Platform (76/100), followed by Beads (74) and Mem0 Memory Layer (72). The top 3 are 4 points apart. Fastest-growing on GitHub over the last 7 days: Claude-Mem (+827.4 stars/day).
| # | Agent | FARS | Trust | Stars | Stars/day (7d) | Latest release | License | Language | Last commit |
|---|---|---|---|---|---|---|---|---|---|
| 1 | Cognee Memory Platform | 76 | 19/29 | 33.1k | +22.0 | v1.6.72026-11-06 |
Apache-2.0 | Python | 2026-11-06 |
| 2 | Beads: Branch-Aware Memory for Coding Agents | 74 | 19/29 | 28.6k | +29.1 | v1.3.42026-11-04 |
MIT | Go | 2026-11-08 |
| 3 | Mem0 Memory Layer | 72 | 16/29 | 68.6k | +57.9 | ts-v3.3.12026-09-25 |
Apache-2.0 | Python | 2026-11-06 |
| 4 | MemPalace Memory | 72 | 18/29 | 60.5k | +35.0 | v3.11.02026-10-14 |
MIT | Python | 2026-11-08 |
| 5 | MemSearch Memory Layer | 69 | 18/29 | 2.8k | +3.1 | v0.4.232026-10-27 |
MIT | Python | 2026-11-06 |
| 6 | Hindsight | 68 | 13/29 | 59.9k | +317.3 | v1.2.02026-11-03 |
MIT | Python | 2026-11-08 |
| 7 | MCP Memory Service | 68 | 18/29 | 2.1k | +5.6 | v12.7.02026-11-07 |
Apache-2.0 | Python | 2026-11-08 |
| 8 | OpenMemory Cognitive Memory Engine | 65 | 16/29 | 4.6k | +2.0 | v1.3.02025-12-20 |
Apache-2.0 | TypeScript | 2026-10-18 |
| 9 | Claude-Mem | 64 | 14/29 | 114.8k | +827.4 | v15.12.02026-11-07 |
Apache-2.0 | TypeScript | 2026-11-08 |
| 10 | OpenViking | 59 | 16/29 | 41.8k | +71.6 | langchain-openvik…2026-11-06 |
AGPL-3.0 | Python | 2026-11-08 |
| 11 | memU Personal Memory | 59 | 13/29 | 14.6k | +2.3 | v1.5.12026-03-23 |
NOASSERTION | Python | 2026-11-05 |
| 12 | Graphiti: Temporal Context Graph Engine for AI Agents | 55 | 12/29 | 32.7k | +34.7 | v0.31.22026-10-20 |
Apache-2.0 | Python | 2026-11-08 |
| 13 | Magic Context | 54 | 12/29 | 2.5k | +4.9 | v0.49.432026-11-08 |
MIT | TypeScript | 2026-11-08 |
| 14 | TencentDB Agent Memory | 52 | 11/29 | 30.6k | +45.6 | v2.0.32026-10-26 |
NOASSERTION | TypeScript | 2026-11-04 |
| 15 | Memori | 47 | 11/29 | 17.5k | +26.1 | v3.3.62026-05-28 |
NOASSERTION | Python | 2026-09-18 |
What qualifies: standalone, maintained tools of this kind that people search for by name. Plugins, skill packs and config layers built on top of them are left out and listed on the host tool's profile. The membership list is kept in our public repository, so every change is on record.
The order is the FARS total; stars only break ties, because popularity measures attention rather than whether something holds up. Two limits are worth knowing before you read the order. These are static reviews — source, documentation and repository metadata, with no runtime reproduction — which caps reliability, effectiveness and verifiability and puts the achievable maximum at 77 rather than 100. And a gap of a few points sits inside the noise of that method. Read the table as a shortlist worth investigating, not as a verdict; every score links to its full dimension breakdown and the evidence behind it. Review method →
1. Cognee Memory Platform
Persistent, self-hosted graph and vector memory for AI agents across sessions.
Where it differs
Against a traditional memory stack that separately operates a graph database, vector database, Redis, and relational database, Cognee documents a demo configuration that places graph state, pgvector, session cache, and metadata in one Postgres instance. That reduces the number of services, but the Postgres graph backend is not presented as production-ready; graph-native backends such as Kuzu or Neo4j remain the documented production direction.
2. Beads: Branch-Aware Memory for Coding Agents
Replace messy markdown plans with a Dolt-powered dependency graph for persistent, structured agent memory.
3. Mem0 Memory Layer
Long-term memory storage and retrieval for AI assistants and agents.
Where it differs
The library path is positioned for testing and prototyping. The self-hosted server targets teams running their own infrastructure and includes a dashboard, authentication, and API keys. The cloud platform targets zero-operations production use; the README labels advanced features as fully included there and as “Teasers” for self-hosting.
4. MemPalace Memory
Keep conversations and project material in a local, verbatim memory store and retrieve relevant context with semantic search.
Where it differs
The README identifies ChromaDB as the default and lists sqlite_exact, rust_exact, Milvus, Qdrant, and pgvector as optional backends. It says the project does not provide comparable benchmarks against Mem0, Mastra, Hindsight, Supermemory, or Zep, so it does not support a performance ranking against them.
5. MemSearch Memory Layer
Persistent, cross-platform semantic memory for AI coding workflows.
Where it differs
MemSearch explicitly treats Markdown as editable source data and Milvus as a rebuildable shadow index, rather than making the vector database the primary memory store. Its documented embedding choices include local ONNX, OpenAI, and Ollama, and its Milvus deployment options include Milvus Lite, Zilliz Cloud, and self-hosted Milvus.
6. Hindsight
A memory layer that helps agents retain, retrieve, and reflect on information over time.
Where it differs
The README positions Hindsight against memory approaches based only on RAG or knowledge graphs. Its documented recall path combines vector, keyword, graph, and temporal retrieval, while retention creates entity, relationship, and time-series representations. The supplied material does not provide a reproducible migration procedure or a feature-by-feature comparison with a named product.
7. MCP Memory Service
Self-hosted persistent memory shared across AI assistants and multi-agent pipelines.
Where it differs
Against Mem0, Zep, and a hand-built Redis plus Pinecone stack, the project emphasizes Apache-2.0 licensing, self-hosting, local ONNX embeddings, MCP, 76 REST endpoints, a typed knowledge graph, and automatic consolidation. Its comparison describes Mem0 and Zep as commercial or enterprise offerings and the DIY option as requiring manual integration. Compared with the MCP-native MemPalace, MCP Memory Service adds REST access, a dashboard, OAuth, multi-user operation, Cloudflare synchronization, and graph relationships, but reports lower LongMemEval R@5 figures. The repository also cautions that MemPalace's 96.6% raw-mode result primarily measures ChromaDB with default embeddings and is not directly comparable to this project's default turn-level ingestion.
8. OpenMemory Cognitive Memory Engine
A self-hosted cognitive memory engine for LLMs and agents, offering real long-term memory with a local-first design, Python and Node support.
Where it differs
Compared to memory systems like Mem0, Zep, or Supermemory, OpenMemory emphasizes multi-sector memory and explainability rather than just vector retrieval. The code examples show its difference from LangChain with Pinecone, offering a more lightweight, local-first approach.
9. Claude-Mem
Persistent, searchable project memory for coding agents across sessions.
10. OpenViking
A browsable context database that unifies agent memory, resources, and skills.
Where it differs
The README benchmarks OpenViking against the native memory of OpenClaw, Hermes, and Claude Code. It reports roughly 80–83% LoCoMo accuracy with OpenViking, alongside lower input-token use and query latency; this makes it most relevant when a separate, browsable context layer is preferable to relying only on host-agent memory.
11. memU Personal Memory
A shared, searchable memory layer that turns coding-agent session history into reusable Markdown skills.
Where it differs
Unlike keeping context inside one assistant, memU shares learned session knowledge among configured hosts through one backend. Compared with memU Cloud, self-hosting stores memory in user-configured SQLite or Postgres, though it still requires an embedding provider.
12. Graphiti: Temporal Context Graph Engine for AI Agents
Build real-time knowledge graphs that evolve with every interaction — know what's true now and what was true before.
Where it differs
Graphiti can be compared to GraphRAG: GraphRAG is suited for static document summarization with batch processing and query latency of seconds; Graphiti is designed for dynamic, evolving context with incremental updates, sub-second latency, and explicit temporal tracking with automatic fact invalidation.
13. Magic Context
Persistent context management and cross-session project memory for OpenCode and Pi coding agents.
Where it differs
Magic Context replaces OpenCode's built-in compaction and requires compaction.auto and compaction.prune to be disabled. It cannot run alongside opencode-dcp; for oh-my-opencode, setup can disable the overlapping preemptive-compaction, context-window-monitor, and anthropic-context-window-limit-recovery hooks.
14. TencentDB Agent Memory
A self-hosted team memory hub that turns conversations, documents, and code into governed, reusable assets for AI agents.
Where it differs
The README compares it with Chat History and Standard RAG. Its stated differentiators are distilled Chat Memory and Skills, Wiki plus a link graph for documents, CodeGraph for code, and governance for ownership, version, status, team sharing, Agent loadouts, and ACLs.
15. Memori
Persistent, structured memory from agent execution and conversation.
Where it differs
The README reports 87% overall accuracy and an average of 721 tokens per query on the LoCoMo long-conversation benchmark, and states that Memori outperformed Zep, LangMem, and Mem0. It also claims more than 36x lower context cost than full-context prompting; these comparisons apply only to the described benchmark setting.