Weekly memory radar refresh: 2026-07-20

[Snseam]

Accepted additions

Papers / benchmarks promoted from primary arXiv sources:

• PM-Bench: prospective memory / pending-intention benchmark.
• MemOps: operation-level remember / forget / update / reflect / no-op benchmark.
• Memory as a Controlled Process: adaptive online memory-control policy.
• Experience Memory Graph: graph-backed error-correction memory over trajectories.
• Why Git: git-bound coding-agent memory substrate.
• Bad Memory: persistent memory prompt-injection benchmark for agent memory files/preferences/KBs.
• MemPoison: persistent memory poisoning benchmark with compositional and dormant-trigger risks.

Product notes updated from official sources:

• Google Agent Platform Memory Bank: memory profiles GA, IngestEvents GA, Gemini Embedding 2 similarity-search configuration.
• Oracle AI Agent Memory: 26.6 docs/product page for hybrid search, CRUD/cascade delete, retention policy, context cards, async APIs, custom extraction.
• OpenAI ChatGPT Memory: current release-note / Help Center memory-summary controls; browser fetch succeeded, curl returns 403.
• AWS Bedrock AgentCore Memory: current release notes rechecked; Harness built-in/BYO memory remains GA surface, no post-window memory delta found.

Updated existing entries

• README / README_cn counts: benchmark catalog 21 -> 25; seed paper notes 15 -> 18.
• papers/index.md manual radar additions.
• benchmarks/index.md, benchmarks/claims/claims.yaml, and docs/benchmarks-landscape.md.
• docs/memory-radar-2026-07.md, docs/signals.md, and docs/product-discovery-log.md.
• Product notes for Google, Oracle, OpenAI, and AWS.

Watchlist / adjacent / rejected

Watchlist:

• Speculate with Memory: memory-augmented speculation is cost/latency-adjacent, not core memory evidence yet.
• FARMA / Your Agent's Memories Are Not Its Own: relevant safety direction, but overlaps with MemPoison and trajectory-forensics until full read.
• AgenticSTS: bounded-memory testbed signal; needs code/data review before catalog promotion.

Adjacent / rejected:

• GitHub repos, stars, awesome-list entries, MCP catalogs, and README benchmark numbers remain discovery only.
• Long-context-only, generic agent-framework, prompt-cache, and catalog-only memory signals were not promoted.
• No title/URL source mismatch was promoted.

Source mismatches / caveats

• Oracle stale 26.6 paths/blog slugs were replaced with reachable official docs/product-page URLs.
• OpenAI Help Center returns 403 to shell curl but loaded through browser fetch; memory-summary controls are present on the current page.
• AWS had no new post-window memory capability; the note records a recheck of existing official release surface only.
• Vendor/product claims remain product-behavior evidence, not independent performance evidence.

Verification

• ruby scripts/verify_memory_refresh.rb passed.
  • papers=989 pdfs=534 stubs=988 full_notes=7 seed_notes=18 products=38 product_archives=37 benchmarks=25
• git diff --check passed.
• Conflict-marker scan passed.
• Targeted canonical URL reachability:
  • 200: arXiv 2607.12385, 2607.12893, 2607.13591, 2607.13884, 2607.14390, 2607.14611, 2607.14651.
  • 200: Google release notes, Oracle 26.6 docs/product page, AWS AgentCore release notes.
  • 403 by curl but browser fetch OK: OpenAI ChatGPT release notes.
• Final review pass found no blocking duplicate/evidence-consistency issues after wording fixes.

Known gaps

• New paper/benchmark notes are seed quality; full PDF/code/data/license normalization is still pending.
• No benchmark scores or vendor superiority claims were normalized in this refresh.
• codex / codex-automation labels were not available in this repository during PR creation.