Eleagent

An elephant's memory. An agent's mind.

A self-hosted personal memory system. Built with privacy as its starting point, Eleagent keeps all storage and model inference on your own hardware—no external LLM API calls.

Capture audio and video in a browser, retain the original evidence, organize observations into a graph, and consolidate the day into memories and a diary during a scheduled sleep cycle.

Eleagent is an experimental reference implementation. Its initial hardware profile uses two NVIDIA DGX Spark systems: daytime perception shares the GPUs with a larger model at night.

Warning

Eleagent is experimental. Expose only the authenticated gateway at your AUTH_PUBLIC_URL. Keep every other service on a trusted private network, and never publish it to the internet directly.

How to start

1. Get two NVIDIA DGX Spark systems, each with 128 GB of unified memory.
2. Verify that you can connect to both machines over SSH. On a computer with SSH access to both, open Claude Code or Codex and paste the prompt below:

Help me deploy Eleagent (https://github.com/okou-ai/eleagent) to my DGX Spark cluster.

First, ask me to confirm the SSH targets for both machines. Do not begin deployment until I confirm them.

After I confirm the targets, follow the agent deployment runbook:
https://github.com/okou-ai/eleagent/blob/main/docs/agent-runbook.md

See the agent deployment runbook for the full deployment procedure.

Access and first initialization

Use one stable HTTPS origin for capture, account sign-in and /mcp; configure it with AUTH_PUBLIC_URL and send public ingress to the authenticated gateway, not directly to observer. During first deployment, ask the owner which email to use, then run the local-only init-owner.mjs --email command with graph-api stopped. It generates a random password and writes the initial credentials to a private mode-0600 file; no password appears in command arguments or logs. See first-owner initialization.

Browser capture uses email/password and a secure HttpOnly session cookie. MCP provides six read-only graph/video tools, uses Better Auth OAuth with PKCE and graph:read / media:read, and serves protected same-origin original downloads with GET/HEAD/Range support. Public signup/setup and shared user API keys are disabled. Authorization uses only public clients with PKCE; each grant receives one access token valid for a year, with no refresh tokens. Revoke an application at /auth/account to end its access immediately.

Private SQLite auth state and its stable secret are persisted together; run one graph-api process per database. Internal observer/archive/GPU credentials and Neo4j authentication remain required machine-to-machine protections—never distribute them to users or remove them to bypass login. GRAPH_API_TOKEN is no longer used.

Licensing

Original project code is AGPL-3.0-only unless a file states otherwise. Dependencies retain their own licenses. Model weights are obtained separately and are not covered by this repository's license.

The DFlash2 draft weights in the experimental accelerated recipe use CC BY-NC-ND 4.0. This recipe is not an unrestricted commercial model stack. Downloading weights yourself does not remove their license conditions. The pinned MiniCPM demo application also lacks an explicit source license; resolve its redistribution terms separately. See third-party notices and the model manifest before choosing a deployment.

References

Core research behind Eleagent's design.

1. T. J. Teyler and P. DiScenna, “The hippocampal memory indexing theory,” Behavioral Neuroscience, vol. 100, no. 2, pp. 147–154, 1986.

2. J. L. McClelland, B. L. McNaughton, and R. C. O'Reilly, “Why there are complementary learning systems in the hippocampus and neocortex: Insights from the successes and failures of connectionist models of learning and memory,” Psychological Review, vol. 102, no. 3, pp. 419–457, 1995.

3. P. A. Lewis and S. J. Durrant, “Overlapping memory replay during sleep builds cognitive schemata,” Trends in Cognitive Sciences, vol. 15, no. 8, pp. 343–351, 2011.

4. B. A. Richards and P. W. Frankland, “The persistence and transience of memory,” Neuron, vol. 94, no. 6, pp. 1071–1084, 2017.

5. J. M. Zacks, N. K. Speer, K. M. Swallow, T. S. Braver, and J. R. Reynolds, “Event perception: A mind-brain perspective,” Psychological Bulletin, vol. 133, no. 2, pp. 273–293, 2007.

6. E. K. Miller and J. D. Cohen, “An integrative theory of prefrontal cortex function,” Annual Review of Neuroscience, vol. 24, pp. 167–202, 2001.

7. M. F. Land and M. Hayhoe, “In what ways do eye movements contribute to everyday activities?,” Vision Research, vol. 41, nos. 25–26, pp. 3559–3565, 2001.

8. M. K. Johnson, S. Hashtroudi, and D. S. Lindsay, “Source monitoring,” Psychological Bulletin, vol. 114, no. 1, pp. 3–28, 1993.

Design notes and implementation links

These are functional design influences—not claims of algorithm reproduction, biological equivalence, or validation of Eleagent.

• Source-linked memory [1]. The graph retains selected memories and links; original media stays in the archive, accessible through Sleep tools.
• Day / night integration [2]. Fast online perception feeds slower nightly consolidation.
• Abstraction and forgetting [3, 4]. Sleep consolidates fragments into reusable memories and lets low-value fragments fade, without deleting recordings.
• Semantic events [5]. Sleep groups meaningful occurrences beyond fixed-duration capture buckets.
• Context-guided attention [6, 7]. Online situation context feeds recent observations and recalled memories back into activity-oriented perception.
• Source monitoring [8]. The perception prompt and Sleep evidence rules distinguish observations, reported speech, and interpretation—including device audio versus in-person speech.