hipocampus
Source: github.com/kevin-hs-sohn/hipocampus Language: JavaScript | Status: Active (26 stars)
Drop-in memory harness for AI agents with a 3-tier memory architecture and 5-level compaction tree.
What It Does
hipocampus manages agent session memory over time. Hot memory (~500 lines) is always loaded. Warm memory (daily logs, knowledge base, plans) is read on demand. Cold memory is searched via qmd hybrid search.
The key innovation is the 5-level compaction tree: raw daily logs get compressed into daily → weekly → monthly → root summaries via LLM-driven summarization. A ROOT.md topic index (~100 lines) gives agents O(1) awareness of what they know.
Key Features
- 3-tier memory: Hot (always loaded), Warm (on-demand), Cold (search)
- 5-level compaction tree with LLM-driven summarization
- ROOT.md topic index for constant-time knowledge awareness
- Hybrid search via qmd (BM25 + vector)
- Claude Code plugin marketplace integration
- Pre-compaction hooks for automatic memory preservation
- File-based, no database
Comparison to kb-mcp
| Aspect | hipocampus | kb-mcp |
|---|---|---|
| Primary use | Agent session memory | Curated knowledge base |
| Data model | Daily logs → compacted summaries | Markdown collections indexed for search |
| Search | qmd (BM25 + vector) | memvid-core (BM25 + optional vector) |
| Write pattern | Continuous (daily logs, auto-compaction) | On-demand (kb_write, manual curation) |
| MCP support | No (skill-based) | Yes (stdio transport) |
Relationship: Complementary. hipocampus handles what the agent remembers from sessions; kb-mcp serves what the agent looks up in reference material.
Patterns Worth Adopting
- Compaction tree — the 5-level summarization pattern is relevant for kb-mcp’s future Knowledge Keeper agent
- ROOT.md topic index — a constant-cost “what do I know?” summary
could complement
list_sections