Refresh weekly agent memory radar evidence

[Snseam]

Accepted additions

Paper seed notes

• Retain or Consolidate? Budget-Dependent Operator Selection for Language Agent Memory (2607.17545)
• Mechanistic Attention Guidance for Agent Memory Refinement (2607.17621)
• AttriMem: Attribution-Guided Process Feedback for Agent Memory Learning (2607.21106)
• Beyond Memory Leaderboards: Evaluating Scientific Memory as Budgeted Context Restoration (2607.16848)
• Your Agent's Memories Are Not Its Own: Forged Reasoning Attacks on LLM Agent Memory and Defenses (2607.05029)
• From Memory to Skills: Evidence-Grounded Co-Evolution Governance for Long-Horizon LLM Agents (2607.16621)
• MemTools: A Unified Research Framework for Interoperable Agent Memory (2607.21404)
• Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents (2607.08716)
• OCR-Memory: Optical Context Retrieval for Long-Horizon Agent Memory (ACL 2026)
• Profile-Graph Memory for LLM Agents / MemHop (2607.19359; page dateline 2026-06-01, first seen locally 2026-07)
• MOSAIC: Accurate and Efficient Long-Term Memory for LLM Agents (2607.16211; page dateline 2026-05-15, first seen locally 2026-07)

Product / benchmark additions

• Databricks Managed Agent Memory product note and archive snapshot, labeled beta / product-behavior evidence.
• Memora benchmark seed note plus origin-protocol claims-ledger event.

Updated existing entries

• BEAM source availability: official repo and Hugging Face dataset links added; performance claims remain non-independent.
• AWS AgentCore Memory: Harness GA and memory execution role / policy signal.
• Google Memory Bank: profiles GA, IngestEvents GA, Gemini Embedding 2 support.
• Anthropic Claude memory-store: agent-memory-2026-07-22 beta header semantics.
• OpenAI Projects memory: project-only scope and project-memory control caveat.
• Alibaba Bailian Memory: OpenSearch Agentic Memory family/alias signal with source-mismatch warning.
• Letta: trajectory format as memory-formation signal.
• Oracle AI Agent Memory: affiliated report watchlist caveat.

Watchlist

• Supra cognitive modes.
• Oracle Agent Memory report results, unless later logged as vendor/affiliated evidence.
• Tacitus, Mnemosyne, SimpleMem, and MCP memory catalogs pending source/license/release-health review.
• Cloudflare Think harness pending product-boundary clarification.

Adjacent / rejected

• Microsoft Foundry Local compaction and Cloudflare Code Mode remain adjacent context/tool-execution signals.
• GitHub stars, README benchmark tables, MCP catalogs, awesome/list placement, and third-party product summaries remain discovery-only.
• Generic OpenSearch docs are not used as evidence for Alibaba Bailian Memory Library.

Source mismatches and access caveats

• 2607.19359 and 2607.16211 use arXiv page datelines rather than ID-month inference; notes now record first-seen context explicitly.
• OpenAI Help Center Projects page returns 403 to shell curl but was source-checked as browser-accessible; only product behavior is recorded.
• Letta trajectory URL normalized to https://www.letta.com/blog/trajectory/.

Verification

• ruby scripts/verify_memory_refresh.rb passed: papers=989 pdfs=534 stubs=988 full_notes=7 seed_notes=26 products=39 product_archives=38 benchmarks=22.
• git diff --check --cached passed.
• Conflict marker scan over README/docs/products/benchmarks/papers/impact-reports found no markers.
• Targeted canonical URL reachability: primary arXiv/ACL/GitHub/Hugging Face/Databricks/Google/Anthropic/Letta/Oracle URLs returned 200; OpenAI Help Center returned shell 403 with browser-accessible caveat.
• Final code-reviewer pass completed; BEAM origin metadata and arXiv date ambiguity findings were fixed before this PR.

Known gaps

• No full-paper reads for the new seed notes.
• No full external link crawl.
• No benchmark reruns or score normalization.
• Product updates remain official/vendor behavior evidence, not independent quality evidence.