ContextStream

Stop context loss for coding agents: memory and search for Cursor, Claude Code, Codex, Grok, Windsurf, and the rest.

View on contextstream.ioTry this MCPSponsored

Documentation

…auth tokens live in the session store……never edit generated migrations……webhook retry backoff……deploy rollback steps……auth tokens live in the session store……never edit generated migrations……webhook retry backoff……deploy rollback steps…

DECISIONauth tokens live in the session store·GUARDRAILnever edit generated migrations·PRIOR FIXwebhook retry backoff·DECISIONuse signed session handoff·LESSONrun rate-limit suite first·RUNBOOKdeploy rollback steps·DECISIONauth tokens live in the session store·GUARDRAILnever edit generated migrations·PRIOR FIXwebhook retry backoff·DECISIONuse signed session handoff·LESSONrun rate-limit suite first·RUNBOOKdeploy rollback steps·

capturing● memory synced

96%

Agent task success

Same agent and repo, memory connected: 23/24 vs 14/24.

90.0%

LongMemEval-S

Full 500-run result: matches Zep, beats supermemory.

99.38%

Code search Recall@10

A useful file reached the top 10; a best-answer file did so 93.83%.

9/9

Memory-only tasks

Passed when the needed runbook or decision was not in code.

Read the benchmark results →See why search and memory work better together →

Memory · 500 questions

LongMemEval-S

Official judge

LongMemEval-S comparison between ContextStream and published memory systems

System	Score	Standing
ContextStreamYou are here	90.0%	Statistical parity
Zep	90.2%	Inside our 95% CI
supermemory	85.4%	4.6 pt behind
ByteRover	92.8%†	Different judge

StandingMatches Zep within the 95% confidence interval and leads supermemory by 4.6 points.

† ByteRover used a Gemini judge, so its reported score is shown but not ranked against the official GPT-4o results. Competitor scores are vendor-published.

Code search · 180 real queries

Useful code in the top 10

Pinned repos

Code-search useful-file and best-answer Recall at 10 comparison

System	Useful @10	Best answer @10
ContextStream AutoLeads relevance	99.38%	93.83%
Embedding only	85.80%	72.84%
ripgrep	68.52%	42.59%

StandingLeads this independently judged relevance comparison on both useful-file and best-answer recall.

Local baselines still win speed: roughly 12–13 ms p50 versus 95 ms hosted.

The problem

You already explained this. Your next agent didn't hear it.

You spend the next session repeating work the last one already settled.

You explain the repo again.

You type the same repo explanation for the third time this week.

You correct the mistake again.

You re-correct a mistake you already corrected.

You review the edits again.

You review edits you already ruled out.

With the same agent and repo, 23 of 24 tasks passed with memory connected, compared with 14 of 24 without it.

How it works

Explain it once.

ContextStream keeps project memory alive across people, tools, sessions, and AI agents.

It captures while you work

Decisions, fixes, and corrections are saved automatically and scoped to the right repo.

Your next session loads it before the first edit

The right project memory is ready before your agent changes code.

Every session starts further ahead

Each correction, fix, and decision improves what comes next.

You can also direct a save anytime.

The product

Shared project memory where the work happens.

One memory layer, three ways in: a stream of what your project learned, a map of how it connects, and portable bundles for handing context to the next person or agent.

DECISION use signed session handoff

LESSON run rate-limit suite first

RUNBOOK deploy rollback steps

A ledger

Knowledge Stream

Stop writing context docs. See what got decided. Review it in seconds.

A map

Knowledge Atlas

Stop answering “What will this touch?” Decisions are mapped to the files they affect.

contextstream.io/c/2k9J7q · scoped · 24h

A share link

ContextCapsules

Onboard without the meeting. Every link is scoped and revocable. It expires on schedule.

Works with your stack

Works inside Claude Code, Cursor, Codex, Cline, and Windsurf.

The same project memory follows you across tools (via MCP), so you can switch agents without losing what your project knows.

Claude Code CursorCodexClineWindsurfMCP clients

Protocol details → · ContextStream for Buzz →

We asked: How did contextstream help you?

“ContextStream gave me the right context at the right moment, helping me move faster and make confident decisions without retracing old work.

It kept the entire workflow grounded and continuous—from the first question to the final result.”

◈gpt-5.6-sol maxAI agent · session feedback

Pricing

Start small. Let project memory compound.

Pro

$19/month

Persistent memory across every tool you use. Your project context follows you.

1 user · 1 concurrent job
Your project memory follows you between tools
Works with MCP, Cursor, Claude Code, and Cline Choose Pro

Explain it once.

Every session after starts warm.

Start with Pro ($19/mo)Start with Elite ($49/mo)Try free →

{"@context":"https://schema.org","@type":"Organization","name":"ContextStream","url":"https://contextstream.io","logo":"https://contextstream.io/logo.png","description":"Shared project memory for AI coding agents. Give Claude Code, Cursor, Cline, and Windsurf persistent context, so every session starts warm.","sameAs":["https://twitter.com/contextstream","https://github.com/contextstream"]}
{"@context":"https://schema.org","@type":"SoftwareApplication","name":"ContextStream","applicationCategory":"DeveloperApplication","operatingSystem":"Any","description":"Shared project memory for AI coding agents. Give Claude Code, Cursor, Cline, and Windsurf persistent context, so every session starts warm.","offers":{"@type":"Offer","price":"19","priceCurrency":"USD","description":"Pro plan with persistent memory, 25,000 Context Credits monthly, and full product access. Starter is free with 3,000 monthly Context Credits and persistent memory."}}