# recall vs EverMemOS

EverMemOS, now shipping as EverOS, is an open-source memory runtime: a model splits conversations into episodes and memory cells, consolidates them into scenes and a profile, and at query time a model judges whether what it found is enough. 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 2026 · recall 0.5.1 · [EverMemOS on GitHub](https://github.com/EverMind-AI/EverMemOS)

## At a glance

EverMemOS set out to be an operating system for memory: conversations cut into episodes by a model, each into cells, cells gathered into scenes, a profile on top, and a model at query time deciding whether what came back is enough. Its 1.0 rewrite, EverOS, drops four databases for markdown, SQLite and LanceDB, closer to recall in shape.

recall was built the small way from the start: the chats your coding agents already keep are the memory, indexed with no model, searched in milliseconds, distilled into notes once a night.

| | recall | EverMemOS |
| --- | --- | --- |
| Runs | On your machine | Self-hosted (1.0: markdown, SQLite, LanceDB), or EverOS Cloud |
| Licence | MIT | Apache 2.0 |
| Model calls to store a chat | **0** | **At ingest and at query time**: boundaries, extraction, a sufficiency check |
| A search takes | **11 ms** local, 47 with embeddings | Under 500 ms p95, a vendor post's claim |
| Price | **Free** | Free in beta; $25 a month after; enterprise custom |
| Chat text leaves your machine | **No**, with a local model | **Yes**: to the model behind ingest and query, OpenRouter by default |
| Claude Code, Codex, Cursor | MCP, skill, session hook | MCP server, Claude Code plugin, Cursor |

## Where both were measured

EverMemOS's paper gives 83.0% on LongMemEval with GPT-4.1-mini answering, on about 2.8k tokens a question by a column it does not define, or 6.7k for the full prompt. recall 0.4.1 scores 88.4% with Claude Haiku 5.5 answering and judging, reading ~44k tokens of whole chats. Different models and judges, so the five points are loose; recall's retrieval alone finds an answering session in the top five 98.7% of the time.

LoCoMo is where EverMemOS's claim is strongest, 93.05%, and where the doubt is: a reproduction with the documented stack scored 38.38%, and the issue closed without a cause. An audit also found 6.4% of LoCoMo's answer key wrong and a full-context GPT-4.1-mini scoring 92.62%. recall measures LoCoMo as retrieval, 92.1% with an answering session in the top five, so there is no answer figure to set beside it.

**Fig 1. recall 0.5.1 and EverMemOS** LongMemEval: answers correct and the context handed to the model, as each reports it.

| Measure | System | Figure | Note |
| --- | --- | --- | --- |
| Answers correct | recall 0.4.1 | 88.4% | Claude Haiku 5.5 answering and judging, reading the top 10 chats |
| Answers correct | EverMemOS | 83.0% | its paper, GPT-4.1-mini answering |
| Context a question | EverMemOS · token column | ~2.8k | its paper's figure, which it does not define |
| Context a question | EverMemOS · full prompt | ~6.7k | its paper's Table 8, the prompt with the answer |
| Context a question | recall 0.4.1 · top 10 | ~44k | ten whole chats, scoring 88.4% |
| Context a question | recall 0.4.1 · windows | ~17.5k | matched windows of the top 5, scoring 84.2% |

LongMemEval_S · EverMemOS's paper, GPT-4.1-mini · recall 0.4.1, all 500 questions EverMemOS's token column is undefined in its paper; its Table 8 puts the full prompt at about 6.7k. LoCoMo is left out: EverMemOS reports 93.05% and a reproduction scored 38.38%, and recall measures retrieval there, not answers.

## Cost

EverMemOS spends model calls twice: at ingest, for boundary detection, extraction and consolidation, and at query time, for the sufficiency check, which reran retrieval on 31% of LoCoMo questions in its paper. The runtime is free to self-host with your OpenRouter key; the cloud is free in beta and $25 a month after.

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.

| | recall | EverMemOS |
| --- | --- | --- |
| Model calls to store a chat | **0** | **At ingest and at query time**: boundaries, extraction, a sufficiency check |
| Price | **Free** | Free in beta; $25 a month after; enterprise custom |
| To take in 300 sessions | **$0**, no model calls | Model calls per episode and per query |
| Cost of memory | One nightly run, your model | Your OpenRouter bill, or the cloud plan |

## Latency

EverMemOS's paper measures no latency; a vendor post claims under 500 ms at the 95th percentile for retrieval. A query that fails the sufficiency check runs a second round, with a model call in between.

recall's search is a local index with no model in the path: 11 ms at the median on ~2,800 chats, 47 ms with the embedding model, and 69 ms at the 95th percentile over a LongMemEval history.

| | recall | EverMemOS |
| --- | --- | --- |
| A search takes | **11 ms** local, 47 with embeddings | Under 500 ms p95, a vendor post's claim |
| Wait at write time | **None**: reads saved chats | Boundary detection and extraction by a model |

## Privacy

Self-hosted EverOS keeps its files on your machine and sends conversations to the model behind ingest and query, through OpenRouter by default, a hosted broker; a local model is possible but not the path the docs take. EverOS Cloud states no data region.

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.

| | recall | EverMemOS |
| --- | --- | --- |
| Runs | On your machine | Self-hosted (1.0: markdown, SQLite, LanceDB), or EverOS Cloud |
| Licence | MIT | Apache 2.0 |
| Chat text leaves your machine | **No**, with a local model | **Yes**: to the model behind ingest and query, OpenRouter by default |
| Compliance | Nothing to certify: no service | No region stated for the cloud |
| Telemetry | **None** | Not stated |

## How memory is built

EverMemOS's memory is layered: episodes cut by topic, cells inside them of three kinds, scenes that group cells, a profile on top, and in 1.0 all of it as markdown over SQLite and LanceDB. Retrieval fuses BM25 and dense search, reranks, then asks a model whether the result will do.

recall keeps every chat whole in a full-text index, ranked five ways and fused, and keeps notes apart from it in markdown under git, consolidated rule by rule once a night. There is no model at query time; the agent that called recall is the reader, and it can search again.

| | recall | EverMemOS |
| --- | --- | --- |
| Captured by | Reading the agents' own chat files | Your code or an MCP client sending messages |
| Memory is | Markdown notes in git, plus an index of every chat | Memory cells in scenes, with a profile |
| Retrieval | Full-text index, five rankings fused; optional local embeddings | BM25 and dense, fused, reranked, then a model checks sufficiency |
| You can edit it | **Yes**, and the edit wins | In the files, since 1.0 |

## Which to pick

**Pick EverMemOS if**

- You want a memory runtime with typed cells, scenes and a profile, and a model checking retrievals at query time.
- You are building on its MCP server or plugins and the per-message model calls through OpenRouter suit you.
- You want a cheap hosted option once the beta ends.

**Pick recall if**

- You want your coding agents to remember with no model calls to index or search, and no broker in the path.
- You want figures you can reproduce yourself rather than a disputed leaderboard.
- You want memory as readable markdown in git, from the chats you already have.

## Questions

**Is EverMemOS the same as EverOS?**

EverOS is EverMemOS's 1.0 rewrite, released in June 2026: the same project and repository, now a local-first Python library on markdown, SQLite and LanceDB. The paper and the benchmark figures describe the earlier system.

**Is the LoCoMo score real?**

EverMemOS reports 93.05%. One reproduction with the documented stack scored 38.38% on the full set, others could not settle it, and the issue closed without a cause. An independent audit also found 6.4% of LoCoMo's answers wrong and a full-context GPT-4.1-mini scoring 92.62%, so the benchmark separates little above 90. recall reports LoCoMo as retrieval, 92.1% with an answering session in the top five, reproducible with `recall eval locomo`.

**Which uses fewer tokens?**

EverMemOS, by its own figures: 2.8k tokens a question by an undefined column, or 6.7k for the full prompt, against recall's 44k reading ten whole chats or 17.5k reading matched windows. recall spends those tokens at read time instead of model calls at write time; on coding issues an agent takes in about 7.8k a question.

**Do both support Claude Code and Cursor?**

Yes. EverOS has an MCP server, a Claude Code plugin and others; recall has an MCP server, a skill and session-start hooks, and reads the transcripts the agents already write.

## Sources

- EverMemOS, now EverOS, on GitHub: https://github.com/EverMind-AI/EverMemOS
- EverMemOS: a self-organizing memory operating system, 2026: https://arxiv.org/abs/2601.02163
- EverOS Cloud: https://docs.evermind.ai/cloud/overview
- EverMind: pricing and latency claims: https://evermind.ai/blogs/evermind-vs-mem
- The LoCoMo reproduction, issue 73: https://github.com/EverMind-AI/EverMemOS/issues/73
- LoCoMo audit: https://github.com/dial481/locomo-audit
- recall's chat benchmark runs, with the other systems' figures: https://github.com/alirng/recall/blob/main/docs/research/retrieval-benchmarks.md