[Cipher208]

Two packages, both found by measuring the real bases

1. memory_graph_edges leaked the other layer's content

It filtered the requested layer on the source node only, while returning both endpoints' content:

WHERE {where} AND s.layer = ?          -- t.layer never checked
... t.content as tgt_content

Reproduced on 7fae1f6: layer="user" returned an agent node's content through a user → agent edge, and the agent side leaked user text symmetrically. The user_id half of this asymmetry was closed by F54 (both endpoints); the layer half was missed.

Both branches now require s.layer and t.layer. Same-layer edges are untouched; only cross-layer ones are hidden.

Latent today — the live bases have zero cross-layer edges — and live the moment graph nodes are re-layered: the pending historical migration would create ~36 000 of them in hermes alone. This is why the fix has to land before any node is moved.

Verifier confirmed the reproduction on base and killed all three guards by mutation (target predicate, source predicate, F54 user predicate). Over-filtering probed: same-layer edges survive for out/in/both, both layers symmetric, node_id=0 unscoped branch, limit clamping intact.

Also corrected count_edges' docstring, which claimed its source-only convention was the one memory_graph_edges used. After this change it is not, and on a cross-layer edge the counter will exceed what the tool lists — stated in the docstring rather than silently diverging.

2. The inventory flagged things it should not

Measured on the three live bases with a marker list of the address forms agents use for the owner:

base	candidates before	after	removed
hermes	649	642	7 wiki_page
mimocode	35	20	15 wiki_page
cowagent	27	26	1 wiki_page

Two precision defects:

• The identity/profile node was a candidate in every base. Its content enumerates the address forms, so it matched several markers at once. It is legitimate user-layer knowledge — moving it would delete the owner's own profile from user memory. Exclusions now mark it decision="exclude" and keep it in the manifest so the decision stays reviewable. The policy block carries only version/source/count/sha256, never the patterns.
• wiki_page is a pointer, not prose. Its content is a filesystem path, so it produced candidates with no text evidence. Removed from the structural set.

summary gains reviewable = candidates - excluded, because the confidence buckets count every emitted candidate and a consumer should not have to know to subtract excluded itself.

Result: 642 / 20 / 26 candidates, exactly one excluded per base — the same #5 profile node in each.

Verification

Two blind verifier passes, both FIX_CONFIRMED, plus corrections for what they found:

• the first pass caught that summary.excluded overlapped the confidence buckets — fixed with reviewable;
• it also caught that one exclusion test used a pattern matching no node, so "the word is not in the manifest" was vacuously true. Rewritten to use a matching pattern and to assert the honest contract: the policy is metadata-only, while a matched candidate's reasons legitimately echo excluded:<pattern>.

Regression probes: read-only snapshot + WAL (committed-but-uncheckpointed rows visible, source bytes unchanged), fingerprint, output guards for memory.db/-wal/-shm/-journal/hardlink/symlink, atomic replace, exact PK validation, duplicate and target-collision fields, user-scoped edge evidence, apply still rejected, byte-identical output on two runs.

Still not done

apply is not implemented. It is a separate package with snapshot, fingerprint gate, postconditions and an explicit write approval. No live base has been modified.

Known divergence left open on purpose: after a re-layering, count_edges will count cross-layer edges that memory_graph_edges will not list. Documented, not fixed — it needs a decision about what the statistic should mean.

Gates

pytest -q            2737 passed
mypy . --strict      63 errors          (identical to 7fae1f6)
ruff check / format  clean / 750 files

[coderabbitai]

Important

• Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

Run configuration

• Source mode: Auto
• Repository: Cipher208/a-memory
• Branch: fix/graph-edge-layer-scope
• File: .coderabbit.yaml
• Commit: 1cbd70f
• Layers: Repository YAML
• Review profile: CHILL
• Plan: Advanced
• Run ID: 90162fd2-dac9-4523-b00b-95b7837515e7

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

Share

• X
• Mastodon
• Reddit
• LinkedIn