LoCoMo · 87.8% accuracy at 460 tokens per question

Recall, not retrain.

The memory layer for AI Agents.

Recalld gives your agents long-term memory and sends back only what answers the question: about 460 tokens, not the whole history. Less noise for your model, more room in its context window. Write conversations and documents once; Recalld extracts facts and updates or supersedes older ones when new information arrives. No need to build your own vector database or memory retrieval pipeline.

See the benchmark

Free tier, no card · EU and US regions live · GDPR-compliant

Three primitives. That's the whole API.

Stop bolting a vector database, a re-ranker, and a dedup pipeline together. Recalld is one endpoint for writing memory and two for reading it.

add Stream in turns or documents. Recalld extracts atomic facts, anchors them in time, and reconciles them against what it already knows, updating, superseding, or merging automatically.
recall Hybrid retrieval plus an LLM curation pass. Returns a short, ranked set of only the facts that answer the query.
search Raw dense-vector lookup when you want the full candidate set and your own model does the filtering. No LLM in the loop.

# Write memory once
POST /v1/agents/{id}/memory
{ "content": "We migrated billing to Stripe on Tuesday." }

# Read back only what matters
POST /v1/agents/{id}/memory/recall
{ "query": "What payment provider do we use?" }

→ only what answers it, ~460 tokens
[{ "fact": "Billing runs on Stripe (since Tue)." }]

Curation is the product.

A pile of "relevant" chunks leaves your agent to sort out the mess, burning context and inviting hallucination. Recalld does the judgement for you.

search

raw

Vector similarity on the raw query, no LLM in the loop. Returns the top candidates and lets your own model do the filtering. Fast, cheap, predictable.

Tokens returned

1,199

Accuracy

85.8%

recall

curated

Runs a selection pass over the candidates and returns only the passages that actually answer the query. That pass runs on our infrastructure, not in your context window.

Tokens returned

460

Accuracy

87.8%

Same questions, same judge. Recall scored 87.8% to search's 85.8%, on 2.6× fewer tokens. Anyone can buy accuracy with a bigger context window. The hard part is not needing one.

Both are official MemJudge runs from 1 November 2026, each with its own ingest. The engine is updated regularly, so current results may differ slightly.

On large models and long agent loops, where the same memory is sent again on every call, those saved tokens also come off your model bill. A DIY vector-database stack has costs of its own, too. See the economics →

Measured, not marketed.

Measured on MemJudge, an open benchmark for AI memory, with LoCoMo (all 10 conversations, 1,986 questions, including the adversarial ones that have no answer in the conversation). A judge model reads exactly what Recalld returns for each question and decides whether it answers it. There is no answer model in between to guess, so a gap in memory counts as a miss.

87.8%

Accuracy recall, 1,986 questions

460

Avg tokens returned 672 at the 95th percentile

93.0%

Adversarial knows when there is no answer

74.8%

Multi-hop reasoning answers spanning many sessions

Full breakdown: single-hop 93.1%, adversarial 93.0%, temporal 87.2%, multi-hop 74.8%, open-domain 57.3%. Recall ran with gpt-6-luna; the judge was claude-sonnet-5-5, the same for every question. Recall's selection pass is an LLM call that runs on our side, so it costs us compute, not your context window; the 460 tokens are what lands in your agent. MemJudge Season 1, 1 November 2026. MemJudge is built by Bit Robotics, the company behind Recalld.

Try it on your own data.

A benchmark is a starting point, not proof for your use case. The free tier is enough to load a few of your own conversations and see exactly what recall returns for the questions you care about.

1,986 questions Adversarial questions included Judge reads the context directly Same judge for every question

See the results on MemJudge

Built for production agents.

Time-aware facts

Recalld can record when facts apply, so a past location and a current one remain distinct.

Fact updates

Recalld evaluates new information against stored facts and can update, supersede, or merge them.

Bring your own key

On Pro and Scale, point Recalld at your own model provider. You pay Google directly; we charge only for orchestration.

Active memory windows

Facts stay stored forever but stay searchable on a plan-tuned window. Nothing is ever deleted.

Regional data residency

Pick the EU or US region when you sign up. Your memory lives there, GDPR-compliant by default, and never moves.

Usage you can read

Credit-based billing with a dashboard in plain credits: ingestion, recall, and search broken out. Export to CSV.

Model Context Protocol

Plug memory into any MCP client.

Recalld ships as a native MCP server. Point any MCP-compatible agent, IDE, or chat app at it and your assistant gets long-term memory as first-class tools. No SDK, no glue code, no bespoke integration.

add The same primitives you already know, exposed as MCP tools. Your client calls them directly. Recalld handles extraction, reconciliation, and curation behind the scenes.
recall The model decides when to remember and when to recall. Curated facts flow straight back into the context window, only what the query needs.
scope Every connection is bound to one agent when you authorize it, so memory stays isolated per client and per tenant.

Works with Claude Cursor VS Code Any MCP host

// mcp.json - add Recalld to any client
{
  "mcpServers": {
    "recalld": {
      "url": "https://eu.recalld.ai/mcp"
    }
  }
}

// No key in the file. Your client authorizes over OAuth
// on first use, and the agent is bound to that connection.

// Your agent now has four tools:
memory_add     - write a turn or document to memory
memory_recall  - get back only the facts that matter
memory_search  - raw candidate lookup, you filter
memory_filters - see what kinds, authors and paths exist

Recalld Chat

The memory layer, as a chat.

Recalld Chat is our own assistant built on the engine: it recalls facts instead of replaying transcripts, keeps every conversation's memory scoped, and shows you exactly what each reply recalled and cost.

Scoped memory Rules you confirm Visible credits Web search

Explore Chat →

Simple, credit-based pricing.

Buy credits, spend them per request. Search is ~40× cheaper than curated recall, so most of your traffic costs almost nothing.

Free

15k credits / month

Default model
30-day active memory
60 req/min
Community support

Starter

$10 /mo

200k credits / month

Default model
6-month active memory
120 req/min
Email support · CSV export

Give your agent a memory.

Recall, not retrain. Start free in the EU or US region today.