Comparison overview

Engram sits in a different part of the memory space than automatic memory engines. It optimizes for human ownership, reviewability, and portability.

Engram strengths

Plain Markdown source of truth.
Human approval before durable writes.
Git-native audit history and sync.
Workspace-first, global-fallback memory.
Agent-agnostic: any agent can read Markdown.
Safety layers: schema validation, secret scan, injection scan, hashes, ignore rules.
No required daemon, database, or cloud account.
Import, observe, archive, graph, benchmark, and repair flows support long-term maintenance.

Less automatic than daemon-based memory systems.
Default search is deterministic lexical search; search --semantic adds deterministic local similarity, not embedding-backed semantic search.
Graph vectors are local hashed word vectors, not semantic embeddings.
Contradiction detection is heuristic and advisory.
deduplicate --semantic uses deterministic local similarity, not external embeddings.
Pattern mining, encryption config, and PR workflow assets exist, but full runtime workflows are not wired yet.
The graph depends on generated tags and summaries.

Optional local embedding provider for graph vectors and search.
Better graph diagnostics explaining why a memory routed.
Benchmark fixtures checked into the repo for regression tracking.
Stronger contradiction review workflow combining graph, quality-check, and archive.
More import tests for agentmemory export variants.
Optional external embedding provider for semantic duplicate detection.
Repair workflows that can propose fixes after reporting invalid memory files.