hipocampus: drop-in 3-tier memory for AI agents, 21.6x over no memory
tom_doerr · x · 2026-08-17
GitHub project hipocampus is a drop-in proactive memory harness for AI agents: zero infrastructure, files only, one command to set up, works with Claude Code, OpenCode and OpenClaw.
On the MemAware benchmark (900 implicit context questions over 3 months of conversation history, testing whether the agent proactively surfaces past context the user never asks about):
- No memory: 0.8%; BM25: 2.8%; BM25 + vector: 3.4%
- Hipocampus (tree only): 9.2%; + vector: 17.3%; + vector (10K ROOT): 21.0%
That's 21.6x better than no memory and 5.1x better than search alone, with the tree still recalling cross-domain context on hard questions with zero keyword overlap.
More from coding & agent
- MinerU: open-source document parser turns messy PDFs into LLM-ready Markdown — abhishek__AI · 2026-08-18
- MinerU: open-source tool turning PDFs, images and Office files into clean Markdown/JSON — abhishek__AI · 2026-08-18
- A dev's 4-phase AI workflow: orchestrator agents merge PRs while he watches TV — brandon_galang · 2026-08-18
- The SKILL.md Fallacy: Coding Agents Should Use Ephemeral Subagents — rseroter · 2026-08-18
- itr8 Integrates New Models Including DeepSeek V4, GPT 5.6 Controls Browser — cantrell · 2026-08-18
- User Reports Claude Cowork Significantly Slower at MCP Tool Use Than Code — evielync · 2026-08-18