Celiums Memory is an open-source cognitive memory engine for AI agents, licensed under Apache-2.0, with all source code public, no open-core split, no paid tiers, and no proprietary core held back. It is positioned as an "engine" rather than an "application": there is no UI and no dashboard to log into; instead, it is embedded into other software as a memory, logging, ethics, and knowledge foundation, such as agents, assistants, tools, and pipelines. It communicates using MCP (Model Context Protocol), so any compatible client (such as Claude Code, Cursor, Continue, Cline, OpenCode, or a custom client) can gain persistent memory, first-person logs, an auditable ethics engine, and per-user circadian rhythms without having to build these capabilities itself.
Main capabilities include: the memory module supports remember / recall, using hybrid retrieval (vector + full-text + emotion/PAD resonance), and has importance scoring, consolidation, lifecycle decay, and circadian rhythm/interoception regulation; the logging module is an append-only, hash-chained first-person agent log, supporting causal chains, arcs, introspection, dialogue, and chain verification; the ethics engine is a four-layer evaluator (A deterministic lexicon, B probabilistic CVaR with categorical CBRN hard blocking, C philosophical scaffolding, K precedent consultation), fully open and including the corpus; the circadian clock regulates arousal and recall according to each user's circadian rhythm, giving the engine time and state awareness; the knowledge module's forage performs hybrid search over user bring-your-own (BYO) skills/knowledge, while the large curated module corpus belongs to another Celiums project, and forage can run without it; for storage, Hyphae-backed Rust-native persistence is used inside the container, with a Durable Object serving as the Cloudflare control plane and recovery log; for recovery, change requests carry X-Celiums-Operation-Id, the DO persists pending, forwards the original command, stores the terminal receipt, advances the contiguous high-water mark, and replays unresolved tail operations after a container failure; for checkpoints, encrypted Hyphae backups are wrapped as versioned binary artifacts, uploaded to R2, referenced by a durable DO pointer, retain the most recent generations, and are restored only to a new verified tenant generation.
In terms of usage, two official run options are provided: Option A is a native Rust server, native MCP/HThe TTP server listens on :3210; option B is MCP stdio mode, runs locally and requires no external services. When connecting an MCP client, point any MCP client at the HTTP endpoint and provide an API key, for example a Claude Code / Cursor style configuration; after that the client can call tools such as remember, recall, journal_write, forage, ethics_trace, and automatic bootstrapping will automatically load prior context in the first response. The ethics knowledge corpus (Layer K, optional) is not in the git tree, but is provided as a v2.0.0 release asset (ethics_knowledge.jsonl, about 31 MB, embeddings precomputed); to enable Layer K, point OPENSEARCH_URL at your OpenSearch and load it, and the loader will download the release asset, verify the SHA-256, create the index with an exact mapping and bulk index (idempotent, can be run repeatedly; --force rebuilds, --dry-run only verifies without writing). Before that, Layer K cleanly abstains, and A + B are unaffected.
Applicable scenarios include: providing cross-session persistent memory for agents or assistants; AI systems that need auditable ethical decisions; interactive applications that need per-user time and state awareness; and product-grade memory foundations that need evidence retention, tenant isolation, and recoverable command semantics.
Regarding limitations, the official statement clearly says: the current dev branch is development work for the Rust and Cloudflare migration, is not yet production-ready, and must not be used for production data, customer traffic, or public release artifacts; the stable user-facing line remains main until the production gates are closed. Production gates include Worker authentication and tenant routing, Durable Object serialization and log durability, container lifecycle, R2 read/write and encrypted recovery, queue delivery and replay, RPO 0 for acknowledged writes, native and container consistency, health canaries, crates.io release, production security review, and disaster recovery drills, etc. Until then, all dev APIs, artifacts, package metadata, and Cloudflare configuration may undergo breaking changes. In addition, the engine itself does not own an LLM or embedding provider; production callers can provide Workers AI, bge-m3, or other compatible models themselves, and during development a deterministic local embedder can also be used.。
