concinno-skills-memory

Progressive-disclosure memory store for LLM agents — three layers (index / summary / archive) so the agent can scan a cheap top layer first and only pay for deeper layers when a keyword hit justifies the cost.

Part of the Concinno ecosystem. Standalone-installable; no runtime dependency on the main concinno package.

Why this exists

Loading the entire conversation history into every prompt is wasteful. The progressive-disclosure pattern keeps the per-turn token cost roughly constant by exposing memories in tiers:

The agent stays at Layer 1 by default and escalates entry-by-entry through expand_memory().

Difference from claude-mem

This package is a clean rewrite of the public progressive-disclosure pattern that claude-mem popularised. We:

• read only the public README / docs to learn the pattern;
• copied no code (claude-mem is AGPL with personal copyright);
• shipped under AGPL-3.0-or-later (the same family Concinno uses).

If you want the full claude-mem MCP server experience, install claude-mem directly. If you want a tiny Python library you can embed in any agent loop and wire into Concinno's ZIQ outcome bus, install this.

Install

pip install concinno-skills-memory

Usage

from concinno_skills_memory import ProgressiveMemoryStore

store = ProgressiveMemoryStore()

# Write — all three layers populated atomically.
eid = store.add_memory(
    topic="releases",
    content="v0.1.0 shipped on 2026-04-28; first public progressive "
            "disclosure release for the Concinno ecosystem.",
    tags=["release", "0.1.0"],
)

# Step 1 — cheap scan.
hits = store.query_memory("release", layer=1)
for hit in hits:
    print(hit.render())   # "releases: v0.1.0 shipped on..."

# Step 2 — escalate the one entry that looks relevant.
summary = store.expand_memory(hits[0].entry_id, layer=2)
print(summary.summary)

# Step 3 — only if L2 is still not enough.
archive = store.expand_memory(hits[0].entry_id, layer=3)
print(archive.content)

ZIQ noise-filter outcome wiring

This package exports a typed callback contract that the Concinno ziq_outcome_bus uses to learn which layer to escalate to per query class. We do not import concinno here — the wire is one-way: Concinno reads our exported names and registers them.

from concinno_skills_memory import (
    NOISE_FILTER_OUTCOME_NAME,
    NoiseFilterCallback,
    NoiseFilterOutcome,
    reference_noise_filter,
)

• NoiseFilterCallback — runtime-checkable Protocol; signature is (query: str, fetched_layer: int, fetched_relevance: float) -> NoiseFilterOutcome.
• NoiseFilterOutcome — float subclass clamped to [0.0, 1.0].
• reference_noise_filter — heuristic implementation that penalises over-fetching (Layer 3 carries a 30 % cost penalty, Layer 2 carries 15 %, Layer 1 carries none).
• NOISE_FILTER_OUTCOME_NAME — stable string the bus keys on ("memory.noise_filter").

Production callers can swap reference_noise_filter for a learned scorer or a Haiku judge. The contract is just the callable shape.

Storage backend

The store is in-memory only by design. Persistence (file, SQLite, vector store, S3) is the caller's choice. A 200-line in-memory core is easier to embed inside any agent loop than a one-size-fits-all backend.

License

AGPL-3.0-or-later. See LICENSE.

Status

Alpha (0.1.0). The public API (three layers, three methods, the outcome contract) is stable for the Concinno 4.4.0 wire. Internal implementation may change before 1.0.