research-coverage: record the 2026 memory-reconstruction literature in the provenance audit (MRAgent, MemHarness, E-mem)

[cdeust]

Three independent 2026 publications converge on the thesis Cortex is built around — that memory is reconstructive, not reproductive. None is cited anywhere in docs/provenance/, CLAUDE.md, or the arXiv drafts. That is a provenance gap in both directions: we lose the strongest available third-party corroboration of the architecture, and we risk presenting as novel what is now independently published prior art.

The papers

Paper	arXiv	Date	Mechanism	Headline result
Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents (MRAgent)	2606.06036	2026-06-04	Cue-Tag-Content graph + LLM-in-the-loop active reconstruction	LoCoMo 84.21 overall (+23.3% rel.), LongMemEval 72.95% (+32% rel.), 118k tokens/sample
MemHarness: Memory Is Reconstructed, Not Replayed	2607.28272	2026-07-30	observation→retrieval→critique→reconstruction→action, GRPO-trained	ALFWorld 85.2% vs 70.1% with raw memory
E-mem: Multi-agent Episodic Context Reconstruction	2601.21714	2026-01-29 (ICML 2026)	Uncompressed contexts + local reasoning per assistant agent	LoCoMo >54% F1, +7.75% over GAM, −70% tokens

Why this is corroboration, not just related work

MRAgent grounds its design in Frankland & Josselyn (2019), Rashid et al. (2016), and Rugg & Renoult (2025) — "retrieval is initiated by contextual cues, which propagate through intermediate representations, progressively reconstructing coherent memory experiences." Cortex already implements that lineage: core/engram.py cites Josselyn & Tonegawa (2020) and core/spreading_activation.py implements Collins & Loftus (1975) cue propagation. MemHarness grounds the same paradigm in Loftus & Palmer (1974) — "human remembering is reconstructive rather than reproductive." E-mem is explicitly "inspired by biological engrams."

Three groups, three independent citation lineages, converging on Cortex's premise within seven months. That is a far stronger claim for the paper and the launch narrative than any single benchmark number.

Scope

• Add a section to docs/provenance/paper-implementation-audit.md covering all three, in the existing per-module format: what each paper claims, which Cortex module implements the equivalent, and whether ours predates or follows it (check git history for dates — do not assert priority without evidence).
• Distinguish honestly per paper: corroborated (we already do it), gap (we do not — see recall: the read path is static retrieve-then-rerank — no evidence-conditioned reconstruction loop (MRAgent, +23% LoCoMo) #361, session_start injects memories verbatim — no applicability critique, and MemHarness measures verbatim replay as actively harmful (76.4% → 70.1%) #363), contradicted (a paper argues against our choice — see falsification risk: E-mem (ICML 2026) argues compression is destructive de-contextualization — test whether Cortex's gist/compress stack costs reasoning fidelity #364).
• Record the shared cognitive-science backbone (Loftus & Palmer 1974; Rugg & Renoult 2025; Frankland & Josselyn 2019) alongside the neuroscience citations already in the audit.
• Do NOT add these to CLAUDE.md benchmark prose until Three different LoCoMo number pairs published across README / docs/benchmarks / CLAUDE.md-referenced baseline — reconcile before arXiv submission #347 is resolved — see the metric-incomparability note filed there.

Acceptance criteria

• All three papers appear in paper-implementation-audit.md with a verdict of corroborated / gap / contradicted per claim, each pointing at a named Cortex module.
• No claim of priority or novelty is made without a dated commit or a published artifact backing it.
• The related-work section of any arXiv draft cites all three.

Sourced from a 2026-08-06 literature sweep; full texts read, not abstracts or blog summaries (§8).

[cdeust]

Extending the coverage scope. A second sweep (2026-08-06) added a security/governance strand that the first three papers do not touch, and it produced four verified gaps rather than architectural aspirations:

A Survey on the Security of Long-Term Memory in LLM Agents: Toward Mnemonic Sovereignty — taxonomizes attacks across six lifecycle phases (write / store / retrieve / execute / share / forget) with the required defense per phase. Its headline findings are directly aimed at systems like ours: "write-gate validation" is "a shared blind spot across every system examined"; "no published architecture covers all nine governance primitives"; "cross-substrate deletion is the frontier." Mapped to Cortex as #365 (write), #367 (store), #368 (retrieve), #366 (forget). The execute-phase (data-plane/control-plane separation) and share-phase (principal-scoped retrieval) classes are not yet filed — they need a scoping pass against the hook layer and the multi-project store before they are actionable.

MemTrace: Tracing and Attributing Errors in LLM Memory Systems — argues "memory errors in LLMs currently lack systematic diagnosis methods" because memory is a black box, and separates retrieval failure from generation misuse of correctly-retrieved content. Their finding that a significant share of failures are retrieval-ranking rather than generation is worth noting against #343 (our harness measures retrieval on synthetic memories) and against the /why receipt path, which establishes presence-in-context but not which stage failed. Not filed separately — the value is diagnostic tooling, and it should be folded into whatever #343 becomes.

Papers surfaced and not yet assessed, for a later pass: MemGym (2605.20833), A-TMA (2607.01935), SimpleMem (2601.02553), TraceMem (2602.09712), ZenBrain (2604.23878), Human-Inspired Memory Architecture (2605.08538), SSGM (2603.11768), MemBench, MemoryAgentBench, Circuit Analysis of agent memory (2605.03354).

Two of those carry priority risk specifically and should be read before the arXiv submission, not after: ZenBrain (a neuroscience-inspired 7-layer memory architecture) and Human-Inspired Memory Architecture both occupy the design space Cortex claims. The audit needs to establish, with dated commits, which mechanisms Cortex implemented first and which are now independently published prior art we must cite.

So this issue's scope becomes: corroborated / gap / contradicted / prior-art-risk — four verdicts, not three.