Context that persists
across sessions and agents
MayaMemo is the reliability layer for AI coding agents — an evidence-backed ledger, a deterministic context compiler, and checkpoints that let any agent resume exactly where the last one stopped.
CLI-first · hosted · works with any MCP-capable agentnpm i -g @mayamemo/cli
# wire memory + checkpoints into agents
mayamemo install opencode claude cursor codex
# switch agents without re-explaining
mayamemo checkpoint create --title "parser"
mayamemo checkpoint resume cp_7f3a --agent codex
Published on npm and PyPI — or install from source in the docs.
Works with the agents you already use
+ any MCP-capable agent — one command wires memory into all of them.
*Same data, same queries, same embedder — reproduce with scripts/mem0-head-to-head.py.
[ what it is ]
The reliability layer, in three parts
Everything a handoff needs — nothing an agent has to re-learn.
Memory
Save decisions, fixes and constraints as searchable, evidence-backed memories. Deterministic retrieval — ranked under a fixed budget, with an explicit reason for every inclusion.
Learn more →Checkpoints
Freeze the working state in one command: summary, next steps, blockers and the compiled context. Pinned, reproducible, diffable.
Learn more →Handoff NEW
Any agent — same tool or a different one — walks in briefed. Zero re-explaining, with measured handoff receipts you can audit.
Learn more →[ 60 seconds ]
The whole loop, in one terminal
Start a session, checkpoint the state, resume with a different agent — that's the entire workflow.
[ proof ]
Proof, not promises
Every claim on this site is a receipt you can rerun. These are the numbers the compiler, the ledger and the governance layer produce on the live deployment.
A hard budget, not a vibe
The context compiler ranks every candidate memory under a fixed request budget and records the run. 549 of 1,500 requests used across the full verification suite — receipts included.
Open the public benchmark →[ how it works ]
Four steps, every handoff
Capture
Your agent works normally — the ledger records decisions, files, commits and results via CLI, hooks or MCP.
Compile
For any task, the compiler assembles the right context: ranked, budgeted, explainable — same inputs, same bytes.
Checkpoint
Freeze the state in one command: summary, next steps, blockers, compiled context. Pinned and reproducible.
Resume
Any agent — same tool or a different one — walks in briefed. Zero re-explaining, measured handoff receipts.
[ benchmark ]
We tested it against Mem0
40 memories, 10 queries (6 keyword + 4 paraphrase), the same embedder on both sides — a live head-to-head, not a marketing chart.
python scripts/mem0-head-to-head.pyRetrieval ↑ higher is better · latency ↓ lower is betterAll receipts →[ infrastructure ]
Built to be audited
Workspace isolation
Every read and write is authorized server-side per workspace. Foreign project ids return 404 — existence never leaks.
Keys hashed at rest
API keys are stored as SHA-256 digests, shown once at creation, revocable instantly. Repo keys stay git-ignored.
Auditable ledger
Decisions, checkpoints, verifications and handoffs are immutable events. Know what, who, and when — replayable.
Deterministic by construction
No LLM in hashing, parsing, filtering or authorization. Same inputs produce byte-identical context, pinned by sha256.
Stop paying the context tax
Every agent switch that starts from zero is lost time. Bring the memory instead.