Skip to content

Refresh August memory radar evidence - #19

Open
Snseam wants to merge 1 commit into
mainfrom
codex/weekly-memory-radar-2026-08-03
Open

Refresh August memory radar evidence#19
Snseam wants to merge 1 commit into
mainfrom
codex/weekly-memory-radar-2026-08-03

Conversation

@Snseam

@Snseam Snseam commented Aug 3, 2026

Copy link
Copy Markdown
Owner

Accepted additions

Papers / seed notes:

  • Zero-Mem: zero-token memory operations for LLM agents (papers/zero-mem-zero-token-memory-operations.md).
  • Memory Provenance Laundering in LLM Agents (papers/memory-provenance-laundering.md).
  • MemHarness: Memory Is Reconstructed, Not Replayed (papers/memharness-memory-reconstructed-not-replayed.md).
  • Filesystem-Based Memory for LLM Agents (papers/filesystem-based-memory-llm-agents.md).

Benchmarks / origin-protocol seed notes:

  • MemSecBench (benchmarks/memsecbench.md) with memsecbench-origin-2026 claims-ledger event.
  • Setoka (benchmarks/setoka.md) with setoka-origin-2026 claims-ledger event.

Updated existing entries

  • Google Agent Platform Memory Bank: setup/topics/TTL/runtime integration behavior.
  • Mem0: n8n/Zapier workflow integration behavior.
  • Zep: valid-time ordering, opaque page token, and is_null date filter behavior.
  • AWS AgentCore Memory: Bedrock Agents Classic migration / AgentCore runtime memory signal.
  • TencentDB Agent Memory: v2 beta/team-memory signal kept as update/watchlist, with vendor benchmark claims bounded.
  • Hindsight: v0.8.6 list_memory_units ingest-age filtering behavior.

Connected surfaces updated: README / README_cn counts and latest-radar links, docs/README.md, papers/index.md, benchmarks/index.md, benchmarks/claims/claims.yaml, docs/benchmarks-landscape.md, docs/memory-radar-2026-08.md, docs/signals.md, docs/product-discovery-log.md, docs/products-landscape.md, and affected product notes.

Watchlist

  • MemTX / MemTxn / ChronoMem: transactional commit, source-supported updates, and semantic rollback are strong governance candidates for the next full source read.
  • MemLens / MemChain / Know It, Act on It / Beyond Retrieval Analytic Memory: strong paper candidates retained for later expansion.
  • Databricks Managed Agent Memory / Memora: already covered by open PR Refresh weekly agent memory radar evidence #18, so this branch does not duplicate those entries.
  • OWASP Agent Memory Guard / Mnemoverse MCP Memory / AMBIENT: useful discovery signals with primary follow-up URLs, but not promoted without full source read.
  • Letta Memory Filesystem: relevant feature signal, but date was not sufficiently verified in the weekly window.
  • inspeximus / ai-memory-mcp / memgres / sqlite-graph-memory: GitHub-only or vendor/self-claimed benchmark signals.

Adjacent / rejected

  • Microsoft Foundry Local Agentic Retrieval: adjacent; boundary with existing Foundry Agent Service Memory is unclear.
  • HAM-VLN / TransMem / hidden-state memory items: adjacent; closer to embodied navigation, hidden-state/context compression, or model-internal memory than core agent-memory lifecycle.
  • GitHub stars, README benchmark numbers, vendor benchmark rows, MCP/catalog placement: rejected as performance or quality evidence; discovery only.

Source mismatches

No new mismatched canonical title/URL pair was promoted. All accepted paper/benchmark additions use direct arXiv URLs; product updates use official docs/changelog/release URLs. Search-snippet dates were not used as canonical source facts.

Verification

  • ruby scripts/verify_memory_refresh.rb passed: papers=989 pdfs=534 stubs=988 full_notes=7 seed_notes=19 products=38 product_archives=37 benchmarks=23.
  • git diff --cached --check passed before commit.
  • Conflict-marker scan passed.
  • benchmarks/claims/claims.yaml parsed successfully via the repo verifier.
  • Targeted canonical URL checks returned HTTP 200 for accepted arXiv URLs and updated official product/watchlist URLs.
  • Final code-reviewer pass returned PASS; one low consistency note was fixed before commit.

Known gaps

  • Seed notes are not full PDF reads; no reported paper scores or product benchmark claims were promoted.
  • No full external link crawl was run.
  • Open PR Refresh weekly agent memory radar evidence #18 remains separate and unmerged; this branch explicitly avoids duplicating its Databricks/Memora coverage.
  • Several strong August paper candidates remain watchlist to keep this weekly diff bounded and reviewable.

Add bounded seed notes and benchmark-origin records for post-window agent-memory papers while keeping vendor and GitHub signals in their evidence classes. Synchronize README, radar, index, landscape, ledger, and product-note surfaces from the staged refresh.

Constraint: docs/weekly-memory-refresh-runbook.md controls the weekly refresh and requires primary-source promotion plus synchronized connected surfaces.

Rejected: promote every arXiv/GitHub discovery candidate | would widen the diff and overstate unreviewed sources.

Rejected: duplicate PR #18 Databricks/Memora entries | open July 27 PR is still separate from current origin/main.

Confidence: high

Scope-risk: moderate

Directive: Keep GitHub/catalog hits discovery-only unless a primary paper/product source is checked.

Tested: ruby scripts/verify_memory_refresh.rb; git diff --cached --check; conflict-marker scan; targeted canonical URL reachability; final code-reviewer pass.

Not-tested: Full semantic read of every candidate PDF and vendor doc.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant