v0.9.0 — causal knowledge updates + multi-session retrieval

v0.9.0 — causal knowledge updates + multi-session retrieval

Release binary: causal-memory (macOS arm64) — the MCP stdio server + CLI.
Usage: run it bare for the MCP server (dsh-mcp-client / DSH plugin / Claude
Code mcpServers), or causal-memory sleep|resolve-updates|ingest ....

New since v0.3.1

• C7 LLM update-resolver — repeated-decision candidates judged by an LLM;
  falsified lessons soft-invalidated. CLI resolve-updates --apply + wired into
  the sleep cycle as consolidation stage 1.7 (no-LLM degrades silently, judge
  failures keep the edge conservatively).
• Vela-style half-life decay tiers — effective_confidence = conf × 0.5^(age/halflife):
  user_feedback/llm 90d, temporal 7d, rule keeps legacy 0.99/day.
• Multi-session multi-pass retrieval — type-agnostic query decomposition (entities +
  temporal anchors) + full-coverage session expansion + verification loop. LongMemEval
  multi-session 42.9% → 57.9% on the same-codebase baseline (+25.6pp vs the published
  32.3%); temporal +5.3pp. Design: docs/design/multi-session-retrieval-2026-08.md.
• PyO3 Python bindings (crates/causal-memory-py).
• DeepSeek Harness native plugin (dsh-plugin/) — clean-named causal tools on
  ctx.tools + a system-prompt section; portable binary resolution (no hardcoded paths).
• Architecture visualization (docs/architecture.html).

Repo layout

• crates/causal-memory — the library (store v11, hippocampus, retrieval, consolidation)
• crates/causal-memory-cli — MCP server + CLI binaries
• benches/ — CausalEval, LoCoMo, LongMemEval, Memora, AMC harnesses
• dsh-plugin/ — DeepSeek Harness plugin

Verification

• 341/341 tests passing, clippy clean (workspace lib + bins)
• Benchmarks: CausalEval v7 (causal capabilities), LoCoMo, LongMemEval, PersonaMem, Memora

v0.3.1 — Multi-hop causal trace

🎯 What's new

This is the first release with the multi-hop causal trace — the key differentiator from flat memory systems (Mem0/Zep/Letta).

New: trace_cause_chain tool

Single-hop trace_cause finds the immediate decision that caused an outcome. But real debugging is chain-shaped:

service crashed ← OOM ← cache had no TTL ← Redis configured without expiry

trace_cause_chain uses a SQLite recursive CTE to walk backward through the causal graph up to N hops, with chain confidence = product of edge confidences (natural decay for longer chains).

This is structurally impossible for Mem0/Zep/Letta — they don't store causal edges.

Changelog (v0.1.0 → v0.3.1)

Version	Feature
v0.1.0	3 MCP tools (record/search/trace)
v0.2.0	Auto-extractor from session logs
v0.2.1	Outcome-overwrite fix + graded confidence
v0.3.0	LLM judge (DeepSeek/OpenAI compatible)
v0.3.1	Multi-hop causal trace (recursive CTE) + SQL parameterization

Stats

• 4 MCP tools
• 10 unit tests pass
• ~1700 lines Rust
• Zero external runtime deps (single binary + SQLite)

Install

git clone https://github.com/JingxuanC/causal-memory.git
cd causal-memory
cargo build --release
# Binary at target/release/causal-memory

Acknowledgments

• PR #1 by @小贺咯 — multi-hop trace, SQL parameterization, research docs
• Research foundation: agent-teardown/insights/04-16