Skip to content

Latest commit

 

History

History
113 lines (80 loc) · 4.09 KB

icm-infinite-context-memory.md

File metadata and controls

113 lines (80 loc) · 4.09 KB

Resource Evaluation: ICM (Infinite Context Memory)

Date: 2026-03-14 URL: https://github.com/rtk-ai/icm Type: GitHub repository / MCP server Score: 3/5 — Integrated Decision: Added as section after Kairn in guide/ultimate-guide.md (~line 11365)

Summary

ICM is a persistent memory MCP server from the rtk-ai team (same authors as RTK/Rust Token Killer). It provides a dual memory architecture: "Memories" (episodic, configurable decay) and "Memoirs" (permanent knowledge graph with 9 typed relation types). Distributed as a single Rust binary with zero external dependencies, installable via Homebrew.

Key Points

  • Single Rust binary, SQLite, zero deps — Homebrew install
  • Dual architecture: episodic decay + permanent knowledge graph in one tool
  • Hybrid search: BM25 30% + vector similarity 70%
  • Auto-deduplication (>85% similarity blocked)
  • Auto-extraction: pattern hooks, pre-compaction, session-start
  • Supports 14 editors/clients (Claude Code, Cursor, VS Code, Windsurf, Zed, Amp, Cline, etc.)
  • 52 stars, 55 commits as of March 2026
  • License: Source-Available (free for individuals and teams ≤20; enterprise license required above)

Scoring

Criterion Score
Relevance to Claude Code users 4/5
Differentiation from existing content 3/5
Maturity / adoption signal 2/5
License openness 2/5
Overall 3/5

Comparison vs Existing Guide Content

Feature doobidoo Kairn ICM
Language Python Python Rust (single binary)
Install pip pip Homebrew
Episodic decay No Yes (biological) Yes (configurable)
Permanent knowledge graph No Yes Yes (Memoirs)
Auto-extraction No No Yes
License MIT MIT Source-Available

Main differentiator: zero-dependency Rust binary lowers install friction for users who struggle with Python environments. Conceptual overlap with Kairn (knowledge graph + decay) is real but the runtime difference is meaningful.

Benchmarks

All benchmarks below are vendor-reported by rtk-ai — not independently verified.

Storage performance (1000 ops, 384d embeddings):

  • Store (no embeddings): 34.2 µs/op
  • Store (with embeddings): 51.6 µs/op
  • FTS5 full-text search: 46.6 µs/op
  • Vector search (KNN): 590.0 µs/op
  • Hybrid search: 951.1 µs/op

Agent efficiency (Haiku model, multi-session):

  • Session 2: 29% fewer turns, 32% less input context, 17% cost reduction
  • Session 3: 40% fewer turns, 44% less context, 22% cost reduction

Knowledge retention (10 questions):

  • Without ICM: 5%
  • With ICM: 68%

Note: The knowledge retention benchmark uses a sample of 10 questions on Haiku — too narrow for generalization.

Fact-Check

Claim Status Source
Storage: 34.2 µs/op ✅ README benchmarks section
Hybrid search: ~951 µs/op ✅ README benchmarks section
29-40% turn reduction ✅ present / ⚠️ vendor-reported README — rtk-ai self-evaluation
68% vs 5% knowledge retention ✅ present / ⚠️ vendor-reported, n=10 README
Source-Available license, free ≤20 ✅ LICENSE file
9 Memoir relation types ✅ README full list
14 supported clients ✅ icm init documentation
52 stars ✅ GitHub as of 2026-03-14

No hallucinations detected. All figures present in the source README.

License Note

Source-Available license. Free for individuals and teams of up to 20 people. Enterprise license required for organizations above 20 people. Contact: license@rtk.ai

This was flagged in the guide entry with an explicit callout. Teams should verify their org size before deploying.

Integration Location

  • New section: guide/ultimate-guide.md after Kairn (~line 11365), before "MCP Memory Stack: Complementarity Patterns"
  • Comparison matrix updated: ICM column added with Runtime and License rows

Upgrade Trigger

Revisit for 4/5 if:

  • Benchmarks independently verified by community
  • Adoption exceeds 500+ stars with sustained commit activity
  • License changes to MIT/Apache