memgov 0.1.0

Provider-agnostic governance layer for AI agent memory: reuse-aware decay, confidence-gated recall injection, and memory-health metrics. Governs the memory you already have (mem0/Zep/vector DB) instead of storing it.

memgov

Provider-agnostic governance layer for AI agent memory. It doesn't store your memory — it governs the memory you already have (mem0 / Zep / a raw vector DB): decays it by recency × reuse, gates recall injection behind a confidence threshold so stale junk never reaches the prompt, and surfaces memory-health metrics. Local-first (SQLite), zero required deps, BYO store.

The problem

The 2025–2026 memory stores (mem0, Zep, LangMem, Letta) solved write and retrieve. But after a quarter of accumulation, teams hit the second-order problem: memory useful in week 1 is actively harmful in week 12, and nothing ages it out. Stores answer "what do I remember?" — none answer "what should I have forgotten, and what's safe to inject right now?" That's governance, and it's missing.

memgov fills that lane:

• Reuse-aware decay — a tunable recency × reuse score, not a crude global TTL.
• Confidence-gated recall — always search, but only inject when retrieval confidence clears a threshold and the memory isn't stale.
• Memory health — active / fading / stale ratios so silent degradation becomes observable before answers get bad.

Install

pip install memgov        # zero required deps

Python 3.11+.

Quick start

from memgov import Governor, Candidate

gov = Governor()                              # local SQLite sidecar

# register memories your store already holds, keyed by the store's native id
gov.track("mem-1", agent_id="bot")
gov.reinforce("mem-1", agent_id="bot")        # user restated it -> reinforce

gov.recompute(agent_id="bot")                 # recompute decay scores + states

# on recall: your store returns candidates; memgov gates what actually injects
candidates = [Candidate("mem-1", confidence=0.82),
              Candidate("mem-9", confidence=0.31)]   # low-confidence
inject = gov.gate(candidates, agent_id="bot")        # only the worthy ones
print([c.mem_id for c in inject])             # -> ['mem-1']

print(gov.health(agent_id="bot"))             # active/fading/stale snapshot

How it works

• Sidecar metadata keyed by your store's native mem_id (memgov never holds the memory text — it tracks the signals decay needs).
• track / touch / reinforce record usage signals; reinforce (user restated/confirmed) is the strongest keep signal.
• recompute rescans an agent's memories and assigns decay_score (0..1) and a state: active | fading | stale | pinned.
• gate is the core: it receives recall candidates and passes only those with confidence ≥ threshold and not stale. pin protects a memory from ever decaying.

See it (no deps, no API key)

python examples/demo_governance.py

Status

Early MVP (v0.1). Implemented: decay scorer, gate middleware, health metrics, SQLite sidecar, CLI. Planned: store adapters (mem0/Zep), MCP-server option, embedding-confidence helpers, health dashboard.

License

MIT.