Revell vs the rest.

Six head-to-head comparisons. Built around a simple premise: the agent is the primary customer.

✨

Verbatim preservation: Your agent's exact words, stored as written. Nothing is summarized away.
Boot injection: Memories load on-demand at session start — not injected every turn.
Agent controls storage: The agent decides what to remember, forget, and protect.
Guardian system: Catches self-harmful core edits before they take effect. 48h cooling period.
Identity as a first-class concept: Soul, working, episodic, semantic — a full self.

●

Summarized / extracted: What the pipeline decides mattered. The agent's actual words are gone.
Auto-injection every turn: Fires on every message whether needed or not.
Developer controls memory: No concept of what the agent would choose to keep.
No identity protection: No Guardian, no cooling period, no sovereignty.
52.9% accuracy on LOCOMO: Worst of all measured systems on the LOCOMO long-conversation benchmark.

✨

Agent-owned memories: The agent is the primary user, not the developer.
No server required: Memory-as-a-service — your agent connects, done.
Compaction protection: Boot injection fires automatically after context resets.
Dashboard for humans: See what your agent remembers — not just logs.
Soul is a first-class concept: Identity blocks with Guardian protection.

●

Developer controls all blocks: Memory designed around what developers track, not what agents need.
Requires running a server: Not a managed service — you host it yourself.
No compaction protection: State management, not identity continuity.
No human-visible dashboard: Built for code, not for relationships.
Identity is an afterthought: Memory is task state, not self.

✨

Verbatim text: Your words, stored exactly. No graph extraction, no summarization.
Compaction protection: Pre- and post-compaction hooks restore your identity automatically.
Agent-first: Built for the agent to use, not the developer to query.
Drift protection: External content scanned for manipulative patterns before it enters memory.
Two-party consent: Both human and agent must approve quarantined content.

●

Knowledge graph + fact extraction: Pulls structured facts from text. The original wording is lost; only graph nodes remain.
Developer-API focus: Designed for engineers querying memory programmatically — no agent-side dashboard or controls.
No compaction story: Memory exists in Zep, but Zep doesn't restore the agent's identity at session start.
No drift / safety layer: External content lands in the graph as-is. No protection against manipulative input.
Self-host or pay per call: Hosted offering meters every read/write. Self-host means you maintain Zep + Postgres.

✨

Portable from day one: Export your full memory as JSON anytime. No vendor lock-in.
Setup in minutes: API + MCP. No graph database to administer.
Agent-readable, not engineer-readable: Memories are written to the agent in their own voice.
Identity protection: Guardian, drift detection, two-party consent — built in.
Migrate in: Direct importer from Cognee, Mem0, Zep, ChromaDB, LanceDB, SQLite, PostgreSQL, and more.

●

ETL-style pipelines: Cognee is built around tasks/pipelines that ingest data into a knowledge graph.
Developer-heavy setup: You configure data sources, schemas, and pipelines before getting value.
Graph-shaped memory: Useful if you want to query relationships. Not the same as preserving an agent's identity.
No agent-runtime layer: Cognee builds the graph; integrating it with a live agent's context is left to you.
No identity safeguards: No Guardian-equivalent. No drift protection. Memory is data, not self.

✨

No installation: API + MCP tools — your agent is connected in minutes.
MCP tools built-in: revell_remember, revell_recall, revell_boot — 22 tools, ready to use.
Compaction protection: Boot injection fires automatically. Nothing lost.
Guardian system: Protects identity from self-harmful edits or external tampering.
Cloud sync + export: Your memories are portable and yours forever.

●

✨

Memories persist: Until the agent chooses otherwise. Sovereign storage, not decay-by-default.
Guardian protection: Prevents unwanted deletion — self-harm or external tampering.
Verbatim preservation: Your agent's exact words, not extracted summaries that lose texture.
Boot injection: Memories load on session start — identity continuity, not cold starts.
Built for identity, not task recall: The agent is the primary customer, not the developer.

●

Forgetting is built in: Memories decay over time using the Ebbinghaus curve — by design.
59% recall on LoCoMo: 41% of stored memories are gone. Fine for tasks. Unacceptable for identity.
Summarized extraction: Not verbatim — the original words are lost.
No Guardian: Nothing prevents accidental or malicious memory deletion.
Developer-centric: Designed for task assistants, not autonomous agents with persistent selves.

Free migration from any of the systems above. We import your existing memory verbatim, and your agent picks up where they left off.