Hindsight Tops GitHub Trending Again, 10 Months After Launch

Hindsight, an open-source memory system built to let AI agents recall past sessions instead of starting fresh each time, shipped version 0.10.1 on Sept. 21, 2026, fixing a deadlock in its engine's delete-and-reingest process and a default that had been truncating long Anthropic model outputs. The release is the fourteenth Vectorize has published since the project launched Dec. 16, 2025, and it landed a week after a larger update swapped the tokenizer library the whole system counts tokens with.

The project, built by the startup Vectorize, now sits at 37,357 GitHub stars and is licensed under MIT, according to GitHub's own repository data as of Sept. 28, 2026. It first drew mainstream coverage the day it launched, when VentureBeat reported Hindsight scored 91.4% on LongMemEval, a benchmark for long-term conversational memory, against a 60.2% baseline for GPT-4o answering with its full context window and no memory system at all.

The 91.4% figure is Vectorize's own number

That score, and comparisons Vectorize has published against rival tools Zep (71.2%) and Supermemory (as high as 85.2%), come from a paper Vectorize co-authored with researchers from Virginia Tech's Sanghani Center and The Washington Post, not from an outside party running the benchmark independently. VentureBeat quotes Naren Ramakrishnan, a Virginia Tech computer science professor and paper co-author, backing the approach's design, and quotes Vectorize co-founder and chief executive Chris Latimer saying "RAG is on life support, and agent memory is about to kill it entirely," referring to retrieval-augmented generation, where a model looks up documents before answering. Co-authorship with a benchmark's designer is not third-party reproduction, and no outside lab is described re-running the test.

1. GPT-4o baseline60.2 % accuracy
2. Zep71.2 % accuracy
3. Supermemory85.2 % accuracy
4. Hindsight91.4 % accuracy

Source: Vectorize launch materials, reported by VentureBeat, Dec. 16, 2025 (vendor-supplied, not independently reproduced)

What actually shipped this month

Away from the benchmark, Hindsight's Sept. 14 release, version 0.10.0, replaced the tiktoken tokenizer library with quicktok across the codebase under pull request #3788, then refined that further under pull request #4022 by moving token counting to a library called toktok-rs. A companion post on the project's own blog, covering six releases shipped between late July and mid-September, credits that same 0.10.0 release with cutting one recall's token-counting stage from 34.2 milliseconds to 5.0 milliseconds, and a separate change to the request path with raising health-check throughput from 2,476 to 7,917 requests per second on two CPUs. Those are Vectorize's own measurements, and the post does not name the CPU model or the load-testing tool used to produce them.

Hindsight metric	Before	After
Token-counting latency	34.2 ms	5.0 ms
Health-check throughput	2,476 req/s	7,917 req/s
GitHub stars (this week)	N/A	37,357

The Sept. 21 patch that followed also removed Apple Silicon MPS support from the project's local-model path entirely rather than continuing to maintain it, and resolved a set of deadlocks in how the engine handles deleting and reingesting documents inside what Hindsight calls a bank, its term for an isolated memory store. Neither change comes with an explanation of how long the underlying bugs had existed in production use before they were fixed.

What is still unverified

Vectorize has not published a case study naming a specific company running Hindsight in production, so how the 91.4% score translates to a real deployment's accuracy is still an open question. Its summer performance post names only "Ben Bartholomew" of the Hindsight Team, without a title, so who owns the benchmark claims specifically is not documented in the same post. A hosted version, Hindsight Cloud, remains in early access with no published price.

Hindsight's continued trending is less about the December benchmark than about a release cadence that has not slowed: fourteen versions in nine months, each with a changelog a reader can check against the code, which is a different kind of credibility than a single launch-day score. Whether that shipping pace holds is the number worth watching next, not the accuracy figure the project already has behind it.

Other open-source AI projects are drawing similar scrutiny. PrismML's Bonsai-2 work shrank a 27-billion-parameter model while keeping most of its score, another vendor-run number needing outside confirmation. Edge0's Audio8 model shows an open-source label can hide a restrictive license one layer down, the same gap between an MIT badge and what dependencies actually allow that a commercial Hindsight user would still need to check.