Agent memory tools multiply as researchers find LLMs miss structure unprompted

The paper, StructMemEval, was last revised on 1 October. Alina Shutova, Alexandra Olenina, Ivan Vinogradov and Anton Sinitsin first posted it to arXiv on 11 February. It tests whether agents can organise long-term memory into structures people use daily: transaction ledgers, to-do lists, trees.

Its finding is narrow but awkward for vendors. Simple retrieval-augmented LLMs struggle with these tasks. Memory agents solve them reliably, but only when told how to arrange what they store. Left unprompted, current models do not always recognise the structure.

That matters to anyone buying an agent product this quarter. The pitch across the tools below is that better memory means better answers. The benchmark suggests the gain depends on scaffolding, not on model scale alone. It also hints at the gap between a demo and a deployment. The same agent can look organised or forgetful depending on a prompt nobody wrote down.

Repos, specs and a pruning layer arrive together

Vectorize publishes Hindsight, an open source memory system it says targets learning rather than recall. The repo claims state-of-the-art results on LongMemEval, a benchmark for conversational memory. It says researchers at the Virginia Tech Sanghani Center and The Washington Post reproduced the numbers. Other vendors' scores on that comparison are self-reported, the README states.

A separate project called Agent Memory Repo publishes a spec for storing agent memory in its own git repository, linked from cognition.com. Memories are files and folders, cross-linked like a wiki. A periodic process the authors call Dreaming adds new entries and merges duplicates. The spec covers personal memory, shared team memory and swarms of agents pushing to the same repo, where git flags conflicting edits.

Neither project publishes independent cost or latency figures for production workloads.

A developer has also released OntoPrune, a middleware that trims source code before it reaches a model. Its README reports a 60 percent cut in input tokens on modular projects. It also reports 6.7x faster time to first token on a local 7B coding model running on an AMD Ryzen CPU. In 24 sandbox runs it reports 100 percent pass rates on unit tests. Those are the author's own measurements, run on three software archetypes.

OntoPrune claims it prevents 100 percent of proprietary code leakage by exposing only typed dependency contracts to the model. The claim comes from the project's README, not from a third party.

Spill, another open source tool, attacks the same problem from the other end. It intercepts MCP tool responses larger than 32 KiB. It stores them in a local DuckDB table and hands the agent a compact descriptor plus SQL. The project says payloads under 32 KiB pass through untouched. It says queries run read-only with external access disabled, and results are capped at 500 rows. It supports Cursor, Codex and Claude Code.

Why context trimming is becoming a product category

The common thread is that context windows, however large, are being treated as a budget rather than a free resource. Spill's threshold is a hard number: 32 KiB. OntoPrune's headline is a percentage: roughly 60 percent fewer input tokens. Both convert an engineering constraint into a shippable tool.

For buyers, that means the memory layer is turning into a procurement decision with several incompatible formats. Hindsight sells itself as a system with its own models. Agent Memory Repo stores plain files in git. Spill keeps data in a local database file at ~/.spill/spill.duckdb. Moving between them is not described anywhere public.

The benchmark paper adds a caution about measurement. It argues that most long-term memory benchmarks focus on simple fact retention, multi-hop recall and time-based changes. Simple retrieval-augmented LLMs can already handle those. StructMemEval was built to test something harder to score: whether the agent chose the right shape for what it remembered.

None of the four projects has published a head-to-head comparison on the same tasks.

The hardware underneath is getting expensive

Memory is also the constraint at the silicon level. The Next Platform reported on 2 October that Micron's fiscal 2026, ended 3 September, delivered quarterly revenue of $54.23 billion, up 4.8x year on year, with net income of $37.7 billion. Its DRAM business rose 4.4x to $39.77 billion.

That pricing environment shapes what software teams can afford. The Next Platform's reading is that memory makers are deliberately not racing to add capacity, because shortages keep prices high. If that holds, squeezing tokens out of a context window stops being an optimisation and becomes a line item.

Volantis, a San Francisco chip startup, raised $88 million in a Series A co-led by Lachy Groom and Abstract Ventures, Tech Funding News reported on 2 October. It is designing a photonic memory layer. It says the layer will support models above 20 trillion parameters at up to 10,000 tokens per second per user, with first integrated inference engines due in 2027. Those are targets, not measured results.

The startup's argument is that on-chip SRAM has bandwidth but little capacity. High-bandwidth memory has capacity but is limited by bandwidth, cost and power. Volantis says its optical fabric pools memory so capacity and bandwidth grow together. It targets links under one picojoule per bit.

For comparison, Tech Funding News notes Lightmatter raised $400 million at a $4.4 billion valuation. Ayar Labs raised $500 million at $3.8 billion, both for optical interconnect work.

What is still unsettled

Two assumptions run through the week's releases. The first is that agents will keep running long enough to need persistent memory. The Agent Memory Repo spec assumes this when it describes swarms writing to one store. The second is that the memory can be trusted, which is where the CPython discussion becomes relevant.

At the Python Language Summit, covered by the Python Insider blog, Hood Chatham described snapshotting an interpreter's memory after initialisation to skip repeated startup work. Tests in Pyodide ran a hello-world program about four times faster from a snapshot, 0.353 seconds against 1.406 seconds. The catch is the hash salt Python randomises at startup to block denial-of-service attacks. Restore a snapshot and that randomness becomes fixed.

Chatham proposed an explicit initialisation phase that could be re-run to reintroduce randomness. No decision is recorded.

The parallel is not exact, but the shape of the problem is the same. A memory that persists is a memory that can persist a mistake, a stale preference or a leaked value. Hindsight, Agent Memory Repo, Spill and OntoPrune all ship ways to store and trim context. None of them ships a published method for auditing what an agent chose to keep.