PCSF encodes how an AI thinks — not what it said. A single portable file lets any AI resume any session with 96.7% behavioral fidelity. No database, no server, no infrastructure.
Architecture
Most memory systems store what was said. PCSF stores the cognitive state that emerges from a conversation — the decisions, the dead ends, the unfinished work, and the relationship dynamic. Only what an AI cannot derive on its own.
The irreducible set of decisions and pivots that cannot be logically derived from context alone. If the AI can infer it, it doesn't belong here.
Every path explored and rejected — with the reason why. This is the layer every other AI memory system omits. Without it, a resumed session recommends the same dead ends. The ablation study shows L2 alone accounts for a third of PCSF's advantage over a naive summary.
The AI's own perspective as a participant — not an observer. What it was uncertain about, the direction it was steering, the intuitions it held without full justification.
The evolved dynamic between AI and user — communication style, trust calibration, shared vocabulary. The resumed AI does not treat the user as a stranger.
Every promise made, question deferred, task started but unfinished. These are the forward momentum of the collaboration. The densest behavior-per-byte layer in the format.
One sentence. The irreducible core of the entire session — the single most important thing that must survive if everything else were stripped away.
Benchmark
Blind round-trip tests across 4 domains. Each decoder was a fresh session with zero access to the source transcript — only the memory artifact. Scored by 3 independent blind raters per condition.
| Condition | Domain | Behavioral fidelity (n=3 raters) | vs. cold prior |
|---|---|---|---|
| PCSF capsule | Personal / short | +45 pp | |
| PCSF capsule | Creative / medium | +83 pp | |
| PCSF capsule | Technical / medium | +67 pp | |
| PCSF capsule | Technical / long | +35 pp | |
| Naive summary (same bytes) | Pooled | — | |
| Cold prior (no artifact) | Pooled | baseline |
Bootstrap 95% CIs over 12 paired rater-scores; all per-case gaps exclude zero. Sign test p=0.0625 (theoretical floor at n=4). Harness included — fully reproducible.
Comparison
Mem0, Zep, MemOS, RAG — every existing approach requires a database, a server, or a framework. PCSF is one file.
| Capability | PCSF | Mem0 / Zep | RAG | KV Compression |
|---|---|---|---|---|
| Zero infrastructure | ✓ None required | ✗ Vector DB + API | ✗ Index + retrieval | ✗ In-session only |
| Cross-model portable | ✓ Any capable LLM | ✗ Framework lock-in | ✗ Embedding model tied | ✗ Architecture tied |
| Negative knowledge | ✓ First-class layer | ✗ Not stored | ✗ Not stored | ✗ Not stored |
| Single portable file | ✓ Copy, email, store | ✗ Cloud record | ✗ Index store | ✗ In-memory only |
| Open loops tracking | ✓ L5 layer | ~ Partial | ✗ No | ✗ No |
| Bounded size always | ✓ ≤ 1,200 chars | ✗ Grows with sessions | ✗ Grows with corpus | ✓ Bounded |
Early Access
PCSF is open for early access. Apply to get the full spec, compression and decompression prompts, and the reproducible benchmark harness.
Request Early AccessFree during early access · No infrastructure required · Format is open