PCM is the lifecycle and attention core of SkillVault's agent memory: Ebbinghaus decay with a savings effect, pinned guardrails that never fade, spreading activation, and Peripheral Attention Engineering (memory delivered as structured prompt slots).
Development continues in ACM (Arboreal Cognitive Mesh), the architecture that runs in production in MemVault. ACM includes everything in PCM plus forgetting by demotion, recall-time spreading activation, episode context, surprisal and absence detection, and the measured benchmark results. This repository is kept for reference and is read-only.
| Module | Purpose |
|---|---|
src/core/decay.ts |
Strength model: initial strength by importance, exponential decay, reinforcement on access |
src/core/pinned-cache.ts |
In-memory cache for pinned guardrails |
src/core/associative.ts |
Association weights and spreading activation |
src/core/margin-heuristic.ts |
Decides when a neural reranker is needed |
src/core/prompt-builder.ts |
Peripheral Attention Engineering slots ([ASKER CONTEXT], [SITUATIONAL CONTEXT], [ANOMALY FLAGS]) |
src/graph/ |
Kùzu knowledge-graph client and code-graph parser |
git clone https://github.com/anthonylee991/pcm.git
cd pcm
bun install
bun test- PCM Specification, version 2.0: architecture, equations and evaluation (also at skillvault.dev/pcm-spec).
- ACM repository: the maintained library, benchmark results and the synthetic surprisal datasets.
MIT © 2026 Anthony Lee & SkillVault Engineering.