AtticDBInstall

Where we win. Where we don't.

Facts from each product's own docs, seen 31 October 2026. Where a vendor publishes nothing, the cell says so.

Nine memory tools

CriterionAtticDBMem0Zep and GraphitiLettaSupermemoryRedis Agent MemoryCogneeEngramCLAUDE.md files
what it storesTurn history per session and keyed factsExtracted memories and entity links[1]Entities, temporal facts, episodes[2]Memory blocks and Git-backed context[3]Extracted facts and profiles[4]Session events and extracted facts[5]Documents, entities, and graphs[6]Project-aware text memories[7]Instructions and markdown notes[8]
storage it needsLocal disk, single Rust binaryVector database; optional graph store[1]Graphiti: graph DB; Zep Cloud: Konig[2]MemFS: Git-backed context repository[3]Vector-graph database; cloud or local[4]Redis; V0 also needs Docket workers[5]Relational, vector, and graph stores[6]Local SQLite + FTS5, single Go binary[7]Local markdown files[8]
model call on writeNoYes, extraction and embeddings[9]Yes, extraction and embeddings[2]LLM tool calls rewrite memory blocks[3]Yes, extraction and embeddings[4]Background extraction and embeddings[5]Local GLiNER or cloud LLM, plus embeddings[6]No LLM or embedding call[7]Manual files; LLM-written auto memory[8]
acked after fsync (published)Yes, after fdatasyncNot publishedNot publishedNot publishedNot publishedNot publishedNot publishedNot publishedNot published
runs fully offlineYesNot publishedNot publishedNot publishedNot publishedNot publishedLocal extraction supported[6]Local binary, no model calls on write[7]Not published
licenceApache-2.0 or MITApache-2.0[1]Graphiti: Apache-2.0; Zep Cloud: proprietary[2]Letta Code: Apache-2.0[3]MIT[4]V0: Apache-2.0[5]Apache-2.0[6]MIT[7]Claude Code: proprietary[8]
MCP6 built-in toolsHosted MCP and OpenMemory[10]Graphiti reference MCP server[2]Letta Code consumes MCP servers[3]Hosted and open-source MCP[4]V0: stdio/SSE MCP server[11]cognee-mcp server and agent plugins[6]Built-in stdio/HTTP MCP server[7]Not published
semantic or graph recallNo, search by wordsVector search; optional graph memory[1]Temporal knowledge graph[2]Not publishedVector-graph recall[4]Vector embeddings for long-term memory[5]Vector and graph recall[6]Text search with SQLite FTS5[7]Plain-text notes, no embedding database[8]

Pick AtticDB if

Pick something else if

You need semantic or graph recall.
Mem0, Zep
You want a full agent harness.
Letta
You want a mature SQLite tool in Go.
Engram

Sources

  1. Mem0 (what it stores, storage it needs, licence, semantic or graph recall): https://github.com/mem0ai/mem0
  2. Zep and Graphiti (what it stores, storage it needs, model call on write, licence, MCP, semantic or graph recall): https://github.com/getzep/graphiti
  3. Letta (what it stores, storage it needs, model call on write, licence, MCP): https://github.com/letta-ai/letta-code
  4. Supermemory (what it stores, storage it needs, model call on write, licence, MCP, semantic or graph recall): https://github.com/supermemoryai/supermemory
  5. Redis Agent Memory (what it stores, storage it needs, model call on write, licence, semantic or graph recall): https://github.com/redis/agent-memory-server
  6. Cognee (what it stores, storage it needs, model call on write, runs fully offline, licence, MCP, semantic or graph recall): https://github.com/topoteretes/cognee
  7. Engram (what it stores, storage it needs, model call on write, runs fully offline, licence, MCP, semantic or graph recall): https://github.com/Gentleman-Programming/engram
  8. CLAUDE.md files (what it stores, storage it needs, model call on write, licence, semantic or graph recall): https://code.claude.com/docs/en/memory
  9. Mem0 (model call on write): https://docs.mem0.ai/platform/pricing
  10. Mem0 (MCP): https://docs.mem0.ai/platform/mem0-mcp
  11. Redis Agent Memory (MCP): https://github.com/redis/agent-memory-server/blob/main/V0/README.md