etchmem 0.1.2

A pluggable, embeddable reconsolidating skill memory engine for AI agents

etchmem

A lightweight, embeddable memory engine for LLM agents. Agents remember things, recall them later, and the stored knowledge quietly matures through use — no extra plumbing required.

pip install etchmem

---

What it is

etchmem gives your agent a persistent memory that improves over time. You deposit raw observations, retrieve them with natural-language queries, and periodically consolidate — at which point the engine synthesizes scattered signals into compact knowledge articles and rewrites anything that's been recalled against fresh context.

It is not a vector database and not an agent runtime. It sits above ChromaDB, which handles all storage and embedding, and exposes exactly three methods.

---

How it works

Memory is kept in three tiers:

• Relational — raw, append-only records with a TTL. Fresh signal lives here.
• Buffer — a working space for deposits and recall-events, waiting for the next consolidation run.
• Injected — synthesized knowledge articles, the primary search target. Only current knowledge is kept; superseded articles are hard-deleted.

The key idea is reconsolidation: every recall() call emits a recall-event into the buffer. When you later call consolidate(), the engine detects which injected articles were used, checks whether fresh signal changes anything, and rewrites them if needed. Knowledge that gets recalled stays current as a side-effect of being recalled. Knowledge nobody asks about simply rests.

Embeddings are computed locally by ChromaDB's built-in model — no external embedding API. Synthesis (the rewrite step inside consolidate) calls an LLM via OPENAI_API_KEY or ANTHROPIC_API_KEY.

---

API

from etchmem import Engine
engine = Engine()   # persists to .etchmem/ in the current directory

remember(data, hint=None, skill=None, metadata=None)

Deposits a text record into the relational tier. No LLM call — cheap.

• data — the text to store.
• hint — optional float 0–1, your importance signal.
• skill — optional scope name; lets you filter recall by agent skill.
• metadata — arbitrary dict (source URL, tags, …).

recall(query, skill=None, top_k=None, hint=None) → list[SearchResult]

Retrieves relevant knowledge and emits a recall-event for future reconsolidation. Results are blended from injected (primary) and relational (fresh signal), ranked by a composite score.

consolidate(num_records="all", method="LIFO") → dict

Runs the consolidation worker: clusters the buffer, forms new injected articles, reconsolidates recalled articles against fresh context, hard-deletes superseded ones. Returns a summary dict with counts (formed, reconsolidated, dropped, kept, flushed, superseded). Requires an LLM API key.

---

Hello world

import os
from etchmem import Engine

os.environ["ANTHROPIC_API_KEY"] = "sk-ant-..."   # or OPENAI_API_KEY

engine = Engine()

# Add some facts
engine.remember("The Eiffel Tower is 330 metres tall including its antenna.")
engine.remember("The tower was completed in 1889 and was originally intended to be temporary.")
engine.remember("It receives about 7 million visitors per year, making it the world's most visited paid monument.")

# Retrieve relevant knowledge
results = engine.recall("How tall is the Eiffel Tower?")
for r in results:
    print(r.score, r.content)

# Consolidate — synthesizes the raw deposits into a compact knowledge article
summary = engine.consolidate()
print(summary)
# {'formed': 1, 'reconsolidated': 0, 'dropped': 0, 'kept': 0, 'flushed': 3, 'superseded': 0}

# Subsequent recalls now hit the synthesized article,
# and any new signal deposited before the next consolidate()
# will be blended in during reconsolidation.
results = engine.recall("Eiffel Tower visitors")
print(results[0].content)

---

License

MIT