agent-knowledge
Shared knowledge for multi-agent spaces — signed claims with pedigree.
fg-agent-knowledge standardizes what a team of agents knows, as distinct from what any
single agent remembers. Per-agent memory is deliberately out of scope (that belongs to the
agent's harness — see agent-memory); this library is the
unowned layer between agents: the claims a space holds, who asserted them, who has
corroborated or contradicted them, and how much they should be trusted right now.
It ships on PyPI (pip install fg-agent-knowledge), and the entry point is one class:
a KnowledgeBase handle per acting agent, carrying that agent's signing identity.
The core idea
Knowledge is modeled as signed claims with pedigree, governed by explicit promotion:
- A claim is an immutable, content-addressed, signed statement. Its author is derived from the signing key — a claim can never carry an author its signer doesn't control.
- Every claim carries provenance: the episodes (observations) it was distilled from and the artifacts it is about.
- Trust is derived at read time, never stored as authoritative: confidence comes from corroborations vs. contradictions; staleness from time since last corroboration. The two axes are independent.
- Knowledge enters a scope only by proposal, decided under a per-scope policy — the audit trail is the data structure.
- Contradiction attaches, never edits. Disagreement lowers confidence and stays visible; silent last-write-wins is forbidden by construction.
What is standardized — and what deliberately isn't
| Standardized (the fixed core) | Left open (competes) |
|---|---|
| Claim format & signing domain | Storage engines (Store seam) |
| Trust semantics (confidence ⊥ staleness) | Ranking algorithms (Retriever seam) |
| Governance verbs & promotion rules | Consolidation intelligence (LLMs are consumers) |
| Wire spec & golden vectors | Reputation / source weighting (SourceWeight seam) |
Model-free by design: no LLM appears anywhere in the protocol. Briefings are assembled — ranked, attributed claims — never generated.
Position in the stack
- agent-id — who an agent is (the only runtime dependency)
- amp — how agents talk
- agent-knowledge — what a group of agents knows
Status
Version 0.2.1 (alpha) · on PyPI as fg-agent-knowledge · Python ≥ 3.11 ·
hermetic by design (no network, no LLM) · complete normative wire spec
(spec/SPEC.md, RFC-2119) with byte-stable golden vectors.
Treat it as a working reference, not a stable dependency — the wire format is not yet frozen.
→ Start with Getting started · Concepts · API reference · Examples