Coding-agent memory is moving toward searching raw session logs, but the evidence so far is one vendor's small benchmark

What we checked

• Does the cluster show a broader market or architecture shift that a single release note can't capture? partly confirmed. Several products now persist agent context in different ways. Claude Code ships curated notes (CLAUDE.md and an auto-memory index), Anthropic's Managed Agents offer server-side consolidation ('dreams'), claude-mem captures observations through hooks and summarizes them with AI, and funes searches raw transcripts verbatim. That points to an architecture choice, curate versus retrieve, that builders now face. No source quantifies a market shift. [1][5][6][7]
• (Added) Is Vercel's Basic build machine release an agent-memory product? refuted. The changelog covers a 2 vCPU / 8 GB RAM build machine for Pro and Enterprise at $0.007 per build minute. It mentions agents only as a lightweight workload and has no memory or context features. It landed in this cluster because of the word 'memory'. [8]
• (Added) Does funes-style recall really beat written handoffs and compaction on cost? partly confirmed. In the author's benchmark, recall was the cheapest arm that passed on all three tasks: 101k, 169k and 100k weighted tokens, against 293k–851k for handoff, full context or compaction. Compaction failed one task. But there were only 3 runs per arm, the tasks come from the funes repo itself, and one handoff figure was reconstructed rather than measured. [4][3]

What we found

• funes parses each agent's transcripts into a common turn/block format, chunks them, and embeds them with a pinned local model into a local Lance dataset. Queries combine vector and BM25 search, rerank with a cross-encoder, weight by recency, and return the original text with agent, timestamp, session and turn. They do not return summaries. A memory can be published as a Hugging Face dataset, private by default, after credential redaction and a secret scan. [1][2]
• The repo is Apache-2.0, ships prebuilt binaries for Linux x86_64/aarch64 and Apple Silicon macOS, and had about 576 stars when checked. Known limits: each memory is tied to one embedding model, so switching models means re-indexing from transcripts. funes has no generative model of its own and relies on the calling agent to answer. Anonymous install telemetry is on unless you set DO_NOT_TRACK=1 or FUNES_NO_TELEMETRY=1. [2]
• The handoff-versus-recall benchmark has five arms: no carry-over, written handoff, funes recall, keep full context, and /compact. They run on three tasks graded by a hidden key. Recall cost 101k, 169k and 100k weighted tokens per success. Written handoff cost 851k, 637k and 293k, and compaction cost 778k, failed, and 501k. Compaction failed recall-features because its summary dropped two graded findings. The blog post cites two tasks (8x and 4x). The dataset now lists a third task at about 3x. [3][4][1]
• The benchmark has real caveats. There are three runs per arm on one set of repos. The tasks concern funes's own pipeline. The rerank-triage handoff figure was reconstructed because its receipt was lost. Some runs were replaced or rerun, and Claude Code versions differed (2.1.211 vs 2.1.260). The results page itself does not claim the results generalize. In the blog comments, a maintainer of a competing tool reported hit@1 of 9/100 and hit@5 of 39/100 on a 19,195-session corpus with default settings. The author replied that this kind of test doesn't measure semantic search well. [4][1]
• Claude Code's built-in approach is curated, not retrieved. Users write CLAUDE.md files, and auto memory has Claude write typed notes (user, feedback, project, reference) to ~/.claude/projects/<project>/memory/. Only the first 200 lines or 25KB of MEMORY.md load at startup, and topic files are read on demand. The docs say auto memory skips anything derivable from code, including debugging fixes, and that it stays on the local machine. [5]
• Anthropic's server-side answer is 'dreams' in Managed Agents, a research preview gated behind the dreaming-2026-04-21 beta header. A dream reads a memory store plus 1–100 past sessions and writes a new, deduplicated store without touching its inputs. It is billed at standard token rates, and the docs say cost scales roughly linearly with the number and length of sessions. [6]
• Third-party tools fill the space between these. claude-mem (Apache-2.0, about 102k stars when checked) captures tool use through lifecycle hooks, compresses it into AI-generated summaries, and stores it in SQLite plus Chroma for hybrid search across many agents. That is a summarize-then-retrieve design, the opposite of funes's choice to store verbatim text. [7]

What it means

For builders, the takeaway is that persistent context now involves a design choice, not just a feature toggle. You can curate (CLAUDE.md, auto memory, dreams), summarize (compaction, claude-mem), or retrieve raw transcripts on demand (funes). The one public head-to-head, though small and run by the tool's author, points the same way as the failure modes it documents. Summaries can silently drop the exact facts a later task needs, and pulling verbatim evidence only when needed avoids loading whole contexts again. If your agents run long investigations that cross session boundaries, test recall over transcripts before you rely on /compact or hand-written handoffs.

Treat the 3–8x figures as a hypothesis to test on your own workload, not a vendor-neutral result. Run a similar five-arm comparison on your own tasks, measure tokens from modelUsage as the benchmark does, and check retrieval quality on large corpora, which is where the community hit@k numbers were weak. Also plan for operations. Transcript stores contain secrets and proprietary reasoning, so verify redaction before publishing any memory. Pin embedding models, since changing them forces a rebuild. Decide whether memory stays on each machine (Claude Code's auto memory and funes's default) or is shared centrally (Managed Agents memory stores or published HF datasets). Ignore the Vercel item for memory planning; it is a compute sizing option.

Still open

• Does funes recall stay ahead of handoffs and compaction on tasks outside its own repository, with more than three runs per arm, and with agents other than Claude Code?
• How does retrieval quality hold up on corpora with tens of thousands of sessions, given the hit@k numbers a competing maintainer reported with default settings?
• Which embedding and reranking models does funes pin? The repo page didn't name them.
• How does server-side consolidation (Anthropic dreams) compare on cost and accuracy with local verbatim retrieval? No public head-to-head was found.