872b0adb48
Split flat docs/ into user-docs/ (guides, config, troubleshooting) and dev/ (ADRs, architecture, extension guides, proposals). Updated docs/README.md index to reflect new paths.
2.2 KiB
Long-Running Memory Fidelity
The core problem: Every compression loses information. Over enough compressions, summaries drift from reality like a photocopy of a photocopy. The system can't easily tell it's happening because it only sees the current summary, not what was lost.
Multi-Tier Memory with Different Decay Rates
| Tier | Decay Rate | Content | Update Strategy |
|---|---|---|---|
| Manifest | Fast (updates every task) | Current state only, <1000 tokens | Continuous overwrite — no history |
| Decision Log | Never decays (append-only) | Every significant architectural decision + rationale | Never summarized, grows linearly |
| Task Archive | Medium | Compressed task completion records | Available for retrieval, not routinely loaded |
The Critical Mechanism: Periodic Reconciliation
All four models converge on some form of automated audit:
- Claude: Every milestone or N tasks — agent compares manifest against actual codebase
- Gemini: Every N commits, spawn a "History Auditor" agent whose sole job is manifest-vs-code comparison
- GPT: Self-healing summaries with checksums — when source files change, invalidate and regenerate
- Grok: Deterministic "Memory Fidelity Audit" node every 5 checkpoints — samples key invariants, scores drift 0-100, auto-rebuilds if drift >15%
The Golden Rule
Never summarize summaries. Each compression layer regenerates from the one below. The codebase is always the lossless source of truth.
The Most Dangerous Form of Drift
Not factual inaccuracy — the loss of "why." The manifest says "auth uses JWT tokens." Three months ago there was a long discussion about why JWT was chosen over session-based auth. That context is exactly what gets compressed away. The append-only decision log solves this by preserving why indefinitely even as what gets continuously compressed.
Phase Boundary Refresh
For very long projects (weeks/months), rebuild the manifest from scratch at phase boundaries by having the agent read the actual codebase + decision log — rather than carrying forward the old manifest with incremental updates. This is the equivalent of defragmenting a hard drive.