Module memfail 

MemFail-style per-operation fault-isolation harness.

Anchor

MemFail frames a long-running agent’s recall pipeline as a chain of distinct operations — summarize → store → retrieve — and makes the per-operation behaviour the testable unit. A failure observed at the recall surface is decomposed into the single stage responsible for it.

mnemo’s operation seams are the three primitives exposed on MnemoEngine:

• store = MnemoEngine::remember — write a [MemoryRecord] into the StorageBackend and the vector + full-text indices, emit a MemoryWrite event, link the hash chain.
• summarize = MnemoEngine::run_consolidation — cluster episodic records by tag overlap and replace each cluster with a structured [Consolidated from N memories] … semantic bundle (SourceType::Consolidation) plus consolidated_from relations. Closest mnemo analogue to MemFail’s “summarize” stage.
• retrieve = MnemoEngine::recall — score the active bank under the active retrieval mode (hybrid RRF by default), return the top-k.

What this harness is

A set of adversarial probes — one per operation — each engineered so a failed assertion is attributable to exactly one stage. The probes are run in order (store → summarize → retrieve); a downstream probe trusts its upstream peers because their probes already passed in the same run.

The canonical run_stale_context_fixture case demonstrates the attribution shape: write the same fact twice (older write at high importance, newer write at low importance), then recall. The default hybrid ranker returns the older / stale record on top. Store + summarize probes succeed (both records are in storage with correct content hashes; no consolidation has run), so the harness attributes the stale recall to the retrieve stage — exactly the MemFail “isolate the operation” output.

What this harness is NOT

• Not a recall-quality benchmark. The probes target seams, not retrieval-mode quality. Use bench/locomo for quality numbers.
• Not a faithful MemFail reproduction. The arXiv reference is prior-art-only; mnemo’s harness exercises the three primitives that the public MCP surface actually exposes.
• Not a write-side guard. A failed probe surfaces an attribution; gating production writes on it is the caller’s choice.

Structs

AttributionReport

Output of run_stale_context_fixture: the canonical MemFail case. attributed_stage is the single stage the harness blames for the observed failure based on which upstream probes still pass.

ProbeOutcome

Outcome of a single adversarial probe.

StageReport

Result of running every probe in a single stage.

Enums

Stage

Identifier for the three operation stages MemFail decomposes a recall into. The harness reports findings keyed on this enum so callers can pivot dashboards by stage.

Functions

run_retrieve_probes

Run the retrieve adversarial probe set against engine.

run_stale_context_fixture

Canonical MemFail attribution fixture.

run_store_probes

Run the store adversarial probe set against engine.

run_summarize_probes

Run the summarize adversarial probe set against engine.