Recoleta Item Note
Show HN: OpenTimelineEngine – Shared local memory for Claude Code and codex
Open Timeline Engine is a local-first shared memory and control platform for AI coding agents, aiming to let agents like Claude/Codex/Cursor remember real working practices across sessions instead of starting from zero…
Summary
Open Timeline Engine is a local-first shared memory and control platform for AI coding agents, aiming to let agents like Claude/Codex/Cursor remember real working practices across sessions instead of starting from zero each time. It extends "memory retrieval" into "timeline-based behavior cloning, policy constraints, audit trails, and dual-AI collaborative execution."
Problem
- Existing AI coding agents usually cold-start every session, repeatedly forgetting codebase conventions, prior corrections, and user preferences, causing the same mistakes to recur.
- Chat memory or prompts alone are insufficient to support reliable autonomous execution: when the same model both decides and executes, hallucinated plans can be acted on directly, and prompt injection can more easily bypass safeguards.
- For day-to-day software development, users also need local data control, auditability, cross-agent shared context, and safety constraints, not just "the ability to recall some preferences."
Approach
- Use a local-first timeline engine to continuously capture signals from real workflows: CLI, Git, VSCode, browsers, MCP sessions, etc., turning "what happened, why it was decided that way, and what the outcome was" into searchable memory.
- Adopt a dual-AI architecture: the executor agent does the work; the API-side advisor agent reads the timeline, provides suggestions/rewrites/safety gating, and does not directly write events, architecturally separating "action" from "supervision."
- Provide shared or isolated workspace memory, allowing multiple executors such as Claude, Codex, and Cursor to share the same workspace memory or remain isolated by executor; retrieval defaults to user-only, with cross-user/cross-executor expansion only on explicit request.
- Introduce a progressive autonomy mechanism: each round is scored on four dimensions—goal clarity, evidence strength, outcome stability, and classifier certainty—to decide whether to continue quickly, perform cautious retrieval/deeper deliberation, or pause for human confirmation.
- Constrain execution through policy and firewall-style safety controls: ABAC, default blocking for sensitive levels,
check_contextbefore edits, protected-directory blocking, instruction text stripping, audit logs, and redaction before embedding and before response, avoiding reliance on prompt-only security.
Results
- The text does not provide standard academic benchmarks or independent experimental evaluation results, so there are no verifiable SOTA numbers; the current public version is v0.3.0, and the project is explicitly labeled experimental and not production-ready.
- The paper/project claims it can build a 25-dimensional behavioral fingerprint covering 6 categories (such as decision-making style, communication, prioritization, context switching, learning style, and emotional patterns), and classify situations into 12 behavioral categories.
- Autonomous decision confidence is weighted by 4 factors: goal clarity 40%, evidence strength 25%, outcome stability 20%, and classifier certainty 15%; these determine whether to proceed, deliberate, or request human input.
- Retrieval budgets are tightly constrained to ≤120ms total per round and ≤60ms per backend/source; when the pgvector quality score is <0.42 or hit count is <2, it can fall back to Qdrant; an example working-set quality threshold is context_quality_score ≥ 0.72.
- Session takeover state is persisted by
session_idwith a 120-minute timeout; execution instructions can be set to expire in 30–120 seconds; the working set is typically refreshed every 6 rounds, and goal discovery is also retriggered every 6th round or after more than 2 failures. - The project cites an external comparative claim: Stanford research reportedly found that "a 2-hour interview can reach 85% personality-clone accuracy"; the author claims this system can passively approach that goal from real working behavior, but the text does not provide its own experimental accuracy, baseline comparisons, or reproducibility data.
Link
Built with Recoleta
Run your own research radar
Turn arXiv, Hacker News, OpenReview, Hugging Face Daily Papers, and RSS into local Markdown, Obsidian notes, Telegram digests, and a public site.