memory / graph.py — GraphMemory

SQLite-persisted knowledge graph mirrored into NetworkX for spreading-activation retrieval. Temporal triplet versioning handles belief updates.

Storage

SQLite at memory_dir/knowledge_graph.db.
In-memory networkx.MultiDiGraph rebuilt at startup via initialize_graph().

Schema

Column	Type	Notes
subject / predicate / object	TEXT	UNIQUE composite key. Subjects/objects lowercased; predicates uppercased.
timestamp	REAL	Last-touched epoch.
weight	REAL	Reinforcement counter.
valid_from	REAL	Creation time.
valid_until	REAL nullable	NULL until superseded.

Indexes on subject and object.

API

Method	Purpose
`add_triplets(triplets) -> int`	Bulk insert / reinforce. Functional predicates (WORKS_AT, LIVES_IN, DRIVES, …) auto-expire prior values.
`delete_by_target(target)`	Substring match deletion.
`wipe_all()`	Truncate.
`get_recent_triplets(limit)`	Most recent edges.
`get_expired_triplets(subject, limit)`	Superseded facts.
`execute_graph_compression(merges)`	Merge synonym nodes (line 216-266).
`get_neighborhood(words, global_limit)`	Spreading-activation BFS retrieval.

Spreading-activation retrieval

Word-to-node fuzzy match: exact → substring → difflib.get_close_matches.
3-hop BFS from each seed.
Score paths by sum of edge weights.
Format winners as (subject)-[PREDICATE]-> (object) -[P2]-> (target).

Functional vs. non-functional predicates

One-to-one predicates (WORKS_AT, LIVES_IN, DRIVES) auto-expire old values when a new fact arrives. Many-to-many predicates (LIKES, KNOWS, HAS) accumulate.

Concurrency

threading.RLock guards both SQLite and the networkx mirror. initialize_graph() rebuilds the entire mirror from disk.