Backlog — observed real-world failure patterns

A record of long-term-memory failure patterns reported by practitioners or measured in evaluation, mapped to where each sits on the state loop. Not a roadmap. Nothing below is scheduled. The current wedge is item 1 and stays the only wedge until its Phase 2 is finished; the rest are ordered by judgement, to be re-ordered when item 1 is done.

source / observation
      ├── identity        whose is it?                     attribution provenance
      ├── authority       may it override?                  authority provenance
      ├── version / hash  is the source still the source?   lineage validity
      └── lineage
            ↓
         memory           validity interval · confidence · state type
            ↓
       governing state    resolution — the current wedge
            ↓
        agent action
            ↓
          outcome         verified result → back to source

Three kinds of provenance, kept apart

kind	question	measured
attribution	whose memory is this?	LoCoMo speaker-swap (external triage); the only external case where a memory's source affects gold
authority	whose observation may govern state?	Mode A Category A; Mode B Family A — matched by a fixed source ranking (BRQ1, F3)
lineage validity	is the evidence that justified this memory still true?	not yet

Lineage validity is distinct from the memory's own validity window. A memory can be inside [valid_from, valid_to), from a high-authority source, and wrong — because the thing it was derived from changed:

code @ hash A   →   memory "config lives in foo.py"     (no valid_to; nothing in the memory changed)
code @ hash B   →   foo.py no longer exists              the memory's evidence is gone

memory validity ≠ source validity. Reported production shape: a note written three weeks ago about code that has since changed; the agent trusts it because it is shorter and reads like a conclusion. Reported fix: bind each memory to a content hash of its source; when the source changes, flag the memory rather than silently rewriting it. The user-facing version is plainer than any authority value: don't let an agent trust a memory after the source that justified it has changed. Coding agents, docs, config and external systems are exactly the sources that change.

Ordered backlog

1. Current — typed governing-state semantics / global-policy stress test. reliagent-bench PR #5. Does any fixed global merge ordering work across heterogeneous state types? Nothing else is added to it.
2. Next candidate — source-linked invalidation / lineage validity. Test shape: a memory derived from an artefact; the artefact changes; is the memory surfaced as suspect before the agent acts on it? Metric: stale-evidence action rate. Would need a source hash on Source and a read-time check; neither exists. Most concrete item; easiest to explain.
3. Repeated-failure / learned-strategy memory. Mode B BRQ3, inconclusive: the B2-lite tasks solved themselves from context. Needs execution-grounded tasks where the failing strategy is not signposted. Separate design.
4. History and governing state coexisting. Issue #5: replace keeps one governing value and no queryable past; keep_both keeps the past and lets no guard fire. Externally reproduced on 11 LongMemEval + 5 GoodAI items.
5. Compaction / materialisation at scale. Reported: long histories end in compaction problems, not contradiction problems. Nothing in the current evaluation reaches it.
6. Policy re-evaluation diff. Replay restores decisions; a policy change cannot yet show which historical decisions it would have changed.
7. Decision-context-dependent governing state. "The remembered failure does not apply here" (Mode B N-04). One scenario, zero false avoidance.

Rules

• An item enters with a reported failure (one line: who, what) or a measured one (which analysis) — never with a design.
• An item leaves when it becomes a frozen evaluation design with pre-registered predictions, not when it becomes a feature.
• Nothing here changes the current wedge's scope.

Sources

Public thread on the "memory is not a list of facts" write-up (2026-09): bitemporality and supersession; replay under policy evolution; SCD as the mature baseline; end utility as not repeating a known failure; source-hash invalidation when the underlying code changes; "authority, lineage, lifecycle, procedural direction" as the shape of a reasoning cache — lineage named independently by two reporters. Motivation, not evidence.