01 / The landscape
The memory problem got solved.
The trust problem got exposed.
Primitives
First-party memory tools, filesystem-backed managed memory, self-improving curation. Store-and-recall is now infrastructure you rent.
Open-source capture
The community engine crowd owns capture → compress → inject. Sixty-thousand-star defaults exist. That fight is over too.
Governance & trust
Nobody verifies whether a stored memory is true. Nobody scores how much to trust it. Nothing travels across vendors. This is the layer we build.
Every major solution is locked to its ecosystem — and no vendor will ever verify memory against its competitors. Cross-family trust is only credible from a neutral layer.
02 / The collection
Three skills. Any runtime. Governed memory.
FORGE RECALL
Point it at ANY existing memory — hand-rolled, Anthropic's Memory Tool, a claude-mem store — and it maps what's there against the 6-layer trust rubric, scores each layer's health, and surgically closes the gaps.
View on GitHub →FORGE GENESIS
Stands up a complete governed memory system on a fresh agent — wrapping whatever storage primitive is present, so the governance travels even when the backend changes.
View on GitHub →FORGE HIVEMIND
Multiple agents — across runtimes AND model families — share memory units through a registry with conflict resolution, sync verification, and a trust trail on every unit.
View on GitHub →03 / The architecture
One 6-layer model under everything.
where to look — a tiny index, not a warehouse
what happened — automated session hooks
what it means — LLM-powered extraction
what is real — human-owned truth (PARA)
how things connect — backlinks & relations
what to remember now — hot cache injection
Plus an optional seventh discipline — Memory Engineering: formalized memory units, retrieval scoring, progressive disclosure, and voice/vision capture via live multimodal models.
04 / Where this goes
The trust layer, quantified.
FORGE SCORE
A portable, quantified trust signal attached to individual memory units — verifiable by any runtime that reads markdown.
PSZN CONSENSUS
The same memory unit verified across model families — Claude, GPT, Gemini, DeepSeek — before it earns trust. Powered by PSZN.dev.
PRIMITIVE ADAPTERS
First-class audit targets for Anthropic Memory Tool exports and claude-mem stores. Govern what you already run.
Your agent already remembers.
Now make it trustworthy.
MIT licensed · markdown + shell, no vector DB required · OpenClaw, Claude Code, Hermes, Agent Zero, NemoClaw, Paperclip & generic runtimes