Causal memory layer for AI agents — MCP server that records decision→outcome relationships. Survives compaction.
A causal memory layer for AI agents. Records decisions and their outcomes as causal relationships, so agents learn from experience across sessions and survive compaction.
Memory frameworks today (Mem0, Zep, Letta, OpenViking, MemOS) store what happened. causal-memory stores why — the causal link between a decision and its outcome. This is the slice every other memory layer misses.
Every agent has the same problem: after N compactions, it forgets why it made past decisions. It reverts to a state where the same bug gets fixed the same wrong way, the same architecture choice gets re-debated, the same lesson gets relearned.
This happens because causal information is the most fragile type under text compaction. Real-LLM benchmark (using grok-build's production compaction prompt):
Compactions (k) Textual recall Causal-table recall --------- 1 100% 100% 2 85% 100% 3 55% 100% 5 45% 100%
The causal table doesn't decay because it lives outside the agent's context window — compaction cannot touch it. See docs/design.md and the full benchmark writeup.
LoCoMo (1,986 questions, deepseek-chat answerer + judge, frozen protocol): overall 65.0% · cats 1–4 59.4% · evidence hit rate 74.4% (BM25) · adversarial abstention 84.3% — three controlled runs with full methodology and failure analysis in docs/benchmarks/locomo.md. Honest reading: factual chit-chat QA is Mem0/Zep's home turf, not ours — but abstention behavior and per-run engineering deltas are documented openly, and the compaction-survival experiment below is the one this system is designed to win.
From the project README.
Add the radar badge to your README — it shows your project was picked up by MCP Radar and links to this page:
[](https://mcp.liqiwa.com/s/JingxuanC--causal-memory.html)
Desktop automation MCP server — computer use for any AI agent: control screen, windows, mouse/keyboard, and Chrome via Model Context Protocol (stdio)
mixelpixx/KonnectAI-assisted PCB design for KiCAD 10. Native KiCAD plugin — a single Rust binary exposing 171 schematic, layout, routing, design-review, and manufacturing tools to Claude, or the LLM of your choosing
pioner92/symbolpeek-mcpAsk for the symbol, not the file. Token-efficient MCP server giving AI coding agents symbol-level reads of TS/JS, Rust, Python, Java, Go, JSON, and Markdown
hongnoul/hwatuVisual verification browser for AI coding agents: daemon-based WebKitGTK, real rendering, ~13ms window spawn
max-ramas/rms-memory-mcpPersistent, local, cross-IDE memory for AI agents — markdown source of truth, LanceDB-powered semantic search, zero cloud dependency
Goldziher/sceptreEasyOCR's accuracy, Rust's speed and footprint — CRAFT + gen2 CRNN OCR over ONNX, as a library, CLI, and MCP server.
The top new MCP servers of the week, every Monday. No spam, unsubscribe anytime.