Compare
Compare
vsHindsight

Hindsight extracts facts, experiences and beliefs with a model on every retain. recall keeps the chats and searches them.

Hindsight is an open-source memory engine with retain, recall and reflect: a model extracts facts, entities and times on every retain into four memory networks, and retrieval fuses dense, BM25, graph and temporal search through a reranker. recall indexes the chats Claude Code, Codex and Cursor already keep, with no model calls, and writes notes overnight with a model you choose.

Checked 17 Oct 2026recall 0.5.1Hindsight on GitHub ↗

01At a glance

Hindsight is the most complete open-source memory engine of the systems here: four kinds of memory, a retrieval stack that runs four searches at once and reranks them, a reflect step that forms beliefs, and installers for every major coding agent. The machine runs a model on everything it retains.

recall is a small tool with one idea: the chats your coding agents already keep are the memory, so index them with no model, search them in milliseconds, and let a model write notes from them once a night.

Table 1 · recall and HindsightThe short answers; the sections below take each in turn.
recallHindsight
RunsOn your machineSelf-hosted on Postgres with pgvector, or Hindsight Cloud
LicenceMITMIT
Model calls to store a chat01 LLM call per retain in the default mode, plus a reranker on recall
A search takes11 ms local, 47 with embeddingsUnder 100 ms, its site's claim
PriceFreeSelf-host free; cloud credits, $5 free; Enterprise
Chat text leaves your machineNo, with a local modelTo the LLM behind every retain, local or hosted
Claude Code, Codex, CursorMCP, skill, session hookMCP endpoint per bank; installer for Claude Code, Codex, Cursor, Copilot

02Where both were measured

Hindsight's paper gives 91.4% on LongMemEval with Gemini 3 Pro answering; its live board now shows 94.6%, with LoCoMo at 92.0% and BEAM at 73.9% at a million tokens. recall 0.4.1 scores 88.4% with the smaller Claude Haiku 5.5 as both answerer and judge. Different models and judges make the ranking loose, and Hindsight's figures were reproduced independently, which is more than most vendors can say. recall's retrieval alone finds an answering session in the top five 98.7% of the time; Hindsight does not publish a retrieval figure.

On coding work Hindsight has installers but no figures. recall finds the past fix in its top five 80% of the time on SWE-ContextBench with no model calls, and its notes at session start lift an agent from 32% to 75% on long MemoryCode histories.

recall 0.5.1 and HindsightLongMemEval: answers correct as each reports it, and the latency of a retrieval.↑ Higher is better, except latency
Fig 1
88.4%of LongMemEval questions answered correctly, Claude Haiku 5.5 reading recall's top 10 chats
98.7%of questions have an answering session in recall's top 5; Hindsight publishes no retrieval figure
Answers correctLongMemEval_S
  • Hindsight · board94.6%
  • Hindsight · paper91.4%
  • recall 0.4.188.4%
Retrieval latencyone recall call↓ fewer
  • Hindsight<100 ms
  • recall 0.4.1 · keywords11 ms
  • recall 0.4.1 · hybrid47 ms
LongMemEval_S · Hindsight's paper and live board · recall 0.4.1, all 500 questions · latency from Hindsight's site and recall's laptop
Hindsight's paper used Gemini 3 Pro answering and GPT-OSS-120B judging; its board's model is not named. recall used Claude Haiku 5.5 for both. Hindsight's LongMemEval figures were reproduced independently at Virginia Tech and the Washington Post.

03Cost

Hindsight's cost is a model call per retain in its default mode, plus embeddings on the way in and a cross-encoder on the way out. Self-hosted, that is your LLM bill, gpt-4o-mini by default or a local model through llama.cpp. The cloud sells credits, $5 free with a GitHub signup, with no rate card published.

recall spends no model calls to index or search. The nightly notes are its one model cost, with a model you choose: nothing with a local one, covered by a subscription, batch prices on an API key.

Table 2 · CostWhat storing and reading memory costs.
recallHindsight
Model calls to store a chat01 LLM call per retain in the default mode, plus a reranker on recall
PriceFreeSelf-host free; cloud credits, $5 free; Enterprise
To take in 300 sessions$0, no model callsAn LLM call per retain
Cost of memoryOne nightly run, your modelYour LLM bill, or cloud credits

04Latency

Hindsight claims under 100 ms for a recall call, and under 100 ms at the 99th percentile at a hundred million memories on its Memlake storage; a retain waits on an LLM unless you use the chunks mode.

recall's search is 11 ms at the median on ~2,800 chats, 47 ms with the embedding model, on a laptop, and there is no retain: it reads the transcripts the agents already saved.

Table 3 · LatencyWhat an agent waits for.
recallHindsight
A search takes11 ms local, 47 with embeddingsUnder 100 ms, its site's claim
Wait at write timeNone: reads saved chatsAn LLM call, unless the chunks mode is used

05Privacy

Self-hosted Hindsight keeps memories in your Postgres and sends each retain to the LLM you configured, which can be local; its Memory Defense scanner redacts 45 patterns of secrets before storage. The cloud is managed by Vectorize with no region stated; Enterprise offers BYOC or on-prem.

recall keeps the index and the notes under your home folder, reads chats read-only, redacts secrets before indexing, sends no telemetry, and makes no network calls except to the model you chose for the nightly notes. With a local model, nothing leaves.

Table 4 · PrivacyWhere your chat text goes.
recallHindsight
RunsOn your machineSelf-hosted on Postgres with pgvector, or Hindsight Cloud
LicenceMITMIT
Chat text leaves your machineNo, with a local modelTo the LLM behind every retain, local or hosted
ComplianceNothing to certify: no serviceNo region stated for the cloud; BYOC on Enterprise
TelemetryNoneA secrets scanner redacts before storage

06How memory is built

Hindsight's memory is structured: world facts, experiences, observations about entities and beliefs that reflect can revise, linked in a graph with times attached. Retrieval runs semantic, lexical, graph and temporal search in parallel and reranks the union into a token budget.

recall's memory is plainer: an index of every chat, and notes in markdown under git typed as preferences, project facts, decisions, lessons and open threads, consolidated rule by rule once a night. Search ranks each chat five ways and fuses the rankings; on retrieval benchmarks it sits with the best published, without the graph or the reranker, which recall's research found little evidence for at this scale.

Table 5 · How memory is builtWhat is kept, and in what shape.
recallHindsight
Captured byReading the agents' own chat filesYour code or an agent calling retain
Memory isMarkdown notes in git, plus an index of every chatWorld facts, experiences, observations, beliefs, in a graph
RetrievalFull-text index, five rankings fused; optional local embeddingsDense, BM25, graph and temporal, fused, reranked
You can edit itYes, and the edit winsThrough the API

07Which to pick

Pick Hindsight if

  • You want a full memory engine with typed networks, a graph and a reranker, and are building something that will use all of it.
  • You want per-repo project memory set up by an installer across many agent harnesses, and the model call per retain is fine.
  • You want a cloud with enterprise deployment options and independently reproduced benchmark figures.

Pick recall if

  • You want your coding agents to remember with no model calls to index or search and nothing to retain by hand.
  • You want every past chat searchable, including the ones from before you installed anything.
  • You want a tool that reads files rather than a server with a database, and memory you can read and revert.

08Questions

Is Hindsight open source?

Yes, MIT, including the engine and the coding-agent installers, with a managed cloud alongside. recall is MIT too, with no cloud.

Which scores higher on LongMemEval?

Hindsight: 91.4% in its paper and 94.6% on its live board, against recall's 88.4%. Hindsight answered with Gemini 3 Pro and was judged by a different model; recall used Claude Haiku 5.5 for both, a much smaller model. recall's retrieval alone finds an answering session in the top five 98.7% of the time, a figure Hindsight does not publish.

Does Hindsight run a model on every memory?

In its default retain mode, yes: a model extracts facts, entities and times from each retain. A chunks mode stores text without the model. recall never runs a model to index or search; only the nightly notes use one.

Do both support Claude Code and Cursor?

Yes. Hindsight exposes an MCP endpoint per memory bank and an installer for per-repo project memory across Claude Code, Codex, Cursor and Copilot. recall connects through an MCP server, a skill and session-start hooks, and reads the transcripts the agents already write.

SourcesHindsight on GitHub · github.com/vectorize-io/hindsightHindsight: retain modes · hindsight.vectorize.io/developer/retainHindsight: retrieval · hindsight.vectorize.io/developer/retrievalHindsight benchmark board · benchmarks.hindsight.vectorize.ioHindsight: agent memory for long-horizon agents, 2025 · arxiv.org/abs/2512.12818recall's chat benchmark runs, with the other systems' figures · github.com/alirng/recall/blob/main/docs/research/retrieval-benchmarks.md