eml-memory 0.1.0

Typed, append-only memory store for agentic Claude / LLM sessions. Hybrid text + structured-tag query.

eml-memory

Typed, append-only memory store for agentic Claude / LLM sessions.

This is the v0.1 carve-out of the structured agentic-memory system used across the Monogate project ecosystem. It gives any LLM agent a typed, append-only, queryable knowledge store that survives across sessions.

Install

pip install eml-memory

Quick start

$ eml-memory add fact "Buzzard drama: capacitor C7 spec mismatch" \
    --source notes/buzzard.md --confidence high
added m_a1b2c3d4ef (fact)

$ eml-memory query "Buzzard drama"
1 match

  [fact      ] [high  ] m_a1b2c3d4ef
    Buzzard drama: capacitor C7 spec mismatch
    source: notes/buzzard.md

Every entry has a typed schema:

field	meaning
id	content + UTC timestamp hash
type	fact / decision / convention / todo / retraction
content	the main statement
source	optional citation
confidence	high / medium / low
tags	optional list of strings
metadata	optional structured dict

Python API

from eml_memory import MemoryStore

store = MemoryStore.default()                     # ~/.eml-memory/store.jsonl

# write
store.add("fact", "Lean theorems = 50",
          source="data/lean.md", confidence="high")

# read
for hit in store.query("Lean"):
    print(hit.id, hit.content)

# retract
store.retract("m_a1b2c3d4ef", "superseded by 2026-05-01 audit")

The store object is project-agnostic; pass MemoryStore(path=...) to point at a per-repo or per-feature JSONL file.

CLI

eml-memory add TYPE CONTENT [--source X] [--confidence high|medium|low]
                            [--tag T ...] [--project P]
eml-memory query TEXT [--type T] [--project P] [--tag T] [--limit N]
                      [--include-retracted] [--verbose] [--json]
eml-memory list [--type T] ...
eml-memory retract ENTRY_ID REASON
eml-memory dashboard
eml-memory version

Override the store path via $EML_MEMORY_STORE or --store PATH.

Storage format

Append-only JSONL — one entry per line. Retractions are first-class entries pointing at the retracted ID; the original entry is never overwritten. This means:

• You can tail -f the store from any process.
• Concurrent writers converge on the next read.
• Audit log is the source file itself.

Roadmap

This is v0.1 — text + tag query only. Planned:

• v0.2 — caller-supplied embedding hooks for hybrid retrieval.
• v0.3 — sync adapters (Linear / Notion / GitHub issues).
• v1.0 — bundled embedding model + SQLite backend for high-volume use, plus the contradiction-scan / drift-check passes from the monogate-research curator.

License

PROPRIETARY-PRE-RELEASE.