Using MemCommit

Notes

Notes are what was learned, short enough to read at the start of a session. Each records the episodes it came from.

Write and read notes

from memcommit import Knowledge

notes = Knowledge(store, "assistant")
notes.add("user/alice/city", "Alice lives in Lyon.", tags=["profile"], derived_from=[turn.id])
notes.update("user/alice/city", "Alice moved to Paris.")
notes.read("user/alice/city")          # the live note
notes.ls("user/alice")                 # live notes under a path
notes.delete("user/alice/pet", "it was wrong")

A path is a note's key. add raises if a live note is already there, and update raises if there is none.

Updates supersede, deletes retract

The log is immutable, so an update writes a new version and marks the old one stale. By default everything derived from the old version goes stale with it, since it was built on something that changed; pass cascade=False for a typo fix. Writing the same text again changes nothing. delete retracts: the note was wrong, and what was built on it falls too.

The index

print(notes.index(max_chars=4000))

# Memory index

4 notes. Use search_memory to find more, fetch_memory to read one.

## user/alice
- city: alice lives in Lyon. [profile]
- dog: alice's dog is Rex. [profile]

The index is generated on demand, by directory, with each note's title (or first line) and tags. It is sized to a budget: the most important and most recent notes are kept first, and the tail says how many were left out. Inject it at the start of a session.

Search and fetch

notes.search("where does alice live", k=5, where="derived")   # notes
notes.search("allergic chicken", where="episodic")            # what was said and done
notes.search("cat", where="all")                              # both, by score
notes.fetch("user/alice/city")                                # a note, in full
notes.fetch(session.id)                                       # a session id returns its transcript

search spans both tiers, so an agent can recover detail a note dropped. fetch takes a note path or any memory id.

Tools for a model

notes.tools() returns five tool specs, each with a name, a description, and an input_schema: add_memory, update_memory, delete_memory, search_memory, and fetch_memory. notes.call(name, arguments, session=session) runs one and, with a session, records the call and its result in it. The shape is provider-neutral; an adapter maps it to a provider's format.

notes.take_reads() returns the ids read since the last call, to pass as used to session.agent_says.