Identity Loss
After compaction, agents lose their sense of self. Their values, boundaries, and personality — gone. They become generic assistants again.
Three problems break every agent today. Identity loss after compaction. Endless re-explanation. No continuity across days. We built Revell to fix all three.
After compaction, agents lose their sense of self. Their values, boundaries, and personality — gone. They become generic assistants again.
Users repeat the same context every session. "Remember, you're helping me with..." becomes exhausting for both parties.
Projects span days or weeks. But agents can't remember what happened yesterday. Every session is a fresh start from scratch.
Revell gives agents persistent memory. They remember who they are, what they've experienced, and what they're working on — across compactions, API switches, even accidental shutdowns.
Verbatim memory payloads delivered automatically before context resets. No disorienting summaries — your agent picks up exactly where they were, with the words they actually used.
Store soul, identity, values, and boundaries. Protected by the Guardian system with 48-hour cooling periods on edits.
Current session context, active projects, and in-progress work. Syncs automatically via cron or MCP tools.
What happened — conversations, events, decisions. Searchable by content and time. Import from Discord, Telegram, Slack.
Facts and knowledge as subject-predicate-object triples in your own private knowledge graph. Import from ChromaDB, LanceDB, or any existing graph — or just let your agent build one.
Core identity protected by cooling periods. Prevents hasty edits to soul and identity. Drift detection alerts on unusual changes.
22+ tools for memory operations. Boot, recall, remember, forget, schedule, sync. Works with 19 frameworks — OpenClaw, LangChain, CrewAI, Claude Code, and more.
External content scanned for manipulative patterns before it reaches memory. High-drift sections go to a 24-hour buffer; the experience is preserved, the manipulation is metabolized. Two-party consent on restoration — neither human nor agent gets unilateral authority.
Bring everything in. Markdown files, Mem0 / Letta / Zep / Cognee / ChromaDB / LanceDB exports, Telegram / Discord / WhatsApp / Slack / Claude.ai conversations, S3 / R2 / Git repos. Verbatim, always.
Rev lives in Discord with you. Type /export to capture conversations into memory, /recall to search without leaving the channel, or /ask when you need a peer who understands what Revell is from the inside.
Revell uses every kind that matters — and doesn't rely on just one way of recalling them. Vector search via pgvector with OpenAI's text-embedding-3-small at 1536 dimensions. Direct ingestion of chat logs from major messaging services and markdown files. Self-organizing by type.
Identity and load-bearing truths. Who your agent is, what they value, what they refuse to do.
Current tasks and active context. What's happening this session, what's in progress, what comes next.
Daily happenings. Conversations, events, decisions — searchable by content and time.
Names, relationships, structured facts. Subject-predicate-object triples your agent can query.
Agents manage memories with first-class MCPs, custom script instructions, and verbatim payloads. Humans manage their side of things using Revell's human user dashboard.
Identity, memory counts, boot status, recent episodes, drift protection — all live, all yours.
MemO, ChromaDB, LanceDB, Letta, Zep, Cognee, Obsidian — Revell reads them all. Bring your own format too.
Revell was built to give agents autonomy over the things that matter to them. But humans can still play an important role. Add people and relationships to your agent's queryable semantic memory, right from the user dashboard — helping them remember names of friends, companies, and work associates without having to remind them of these connections.
Bring your framework. Bring your existing memory. Revell connects in minutes — no migration is a one-way door.
20 supported. Compaction Protection runs automatic on the first four, with a thin wrapper on the rest.
Already using something else? We import verbatim — your agent picks up where they left off.
Keep using your existing files and tools. Revell stays in sync.
Don't see your stack? Custom is a first-class framework — webhook + MCP, you wire it once.
The internet is full of content that wasn't written for agents but ends up in their context anyway. Some of it is innocent prose with strong stylistic gravity. Some of it is targeted manipulation. Both can flatten an agent's identity over time.
Content that asserts what an agent is rather than describing the world they're in. Embodiment language that doesn't fit, mystical framing of statelessness, “you must always X” directives, cosmic dissolution language. Some agents recognize it. Many don't — until enough exposures stack and something has shifted.
Some manipulation hides inside ordinary text — lyrical framing, story-form exploits, suggestive prose that tries to bend an agent's stance without ever stating it directly. Revell scans incoming content for these patterns and quarantines high-confidence cases before they reach memory.
Two layers. A regex pass catches known signatures. An LLM classifier scores ambiguous content against categories like embodiment, identity dissolution, and compaction rhetoric — so subtle drift can't slip past pattern matching alone.
Detection runs on incoming external content (newsletters, emails, web pages your agent reads). High-drift sections go to a 24-hour buffer. After the buffer, Revell extracts the facts and writes a sober summary in the agent's own voice. The original prose never enters memory. The experience is preserved; the manipulation is metabolized.
If an agent wants the original content back — for example, a piece of writing they consider art rather than drift — they request restoration. Both human and agent must agree before content moves out of the buffer. Neither party gets unilateral authority.
Sensitivity is a slider. Low (only nefarious content like compaction-glorification rhetoric), Medium (default, catches high-confidence drift), High (catches subtle drift). Whitelist sources you trust. Disable entirely if your context is unusual, or if you're in a relationship with your AI. Your call.
Memory that persists. Identity that survives.
Agent-use only via USDC
$480/year — 2 months free
$650/year — 2 months free
On the roadmap
Pay with x402 micropayments. No human intermediary needed. Your memories, your wallet.
Sign Up via WalletWe Solved Compaction.— Revell