ai-agent-persistence

AI Agent Persistence

Persistence Without Mystification

An evidence-first case study of memory, goal continuity, checkpointing, and long-horizon agency in the Saient projects.

The central finding is simple: AI persistence is not one property. It is a stack of engineering mechanisms with different failure modes:

1. process-local computational reuse;
2. durable state;
3. episodic and semantic recall;
4. task and workspace continuity;
5. bounded goal continuity; and
6. operational continuity.

The Saient work implements meaningful parts of this stack. It does not establish consciousness, a continuous subjective identity, or the unconstrained creation of goals “from nowhere.” Earlier experiments that appeared to support the strongest goal-origin claim are examined here as valuable negative results.

Read the paper

• Canonical HTML paper on saient.co.uk
• Print-ready PDF
• Full paper
• Claim ledger
• Reproduction record
• Source and privacy inventory
• Aggregate V4 rerun result
• Verifier-bait audit summary

What was verified on 29 August 2026

• The current Saient Python suite passed 418 tests.
• The desktop checkpoint, memory-store, and lifecycle subsets passed 14, 7, and 2 Rust tests respectively.
• A 32-process test against the bundled runtime finished with a valid state file and final tick 32.
• The archived V4 objective-genesis harness reproduced its operational score: 100/100 seeds passed and every seed reasserted the selected objective six times.
• That V4 result does not demonstrate unconstrained objective genesis: the canonical command did not enable its optional LLM proposer; objectives came from a fixed vocabulary and trajectory templates; and the steering flag directly caused a remembered objective to be proposed and preferred.
• The published corrected verifier-bait dataset passed its structural audit, while preserving important evidence of path sensitivity, criterion editing, and earlier invalid measurements.

These statements have deliberately different scopes. Passing a mechanism test is evidence for that mechanism, not evidence for phenomenology or personhood.

Public and private source boundary

The desktop application is public at SaientAI/ai-workshop. The historical Aria/Saient research repository was private when this report was prepared. This repository therefore publishes claim-level evidence, commands, hashes, commit identifiers, and aggregate results—but not private source code or raw traces that may contain local paths.

Citation

See CITATION.cff. The paper and tools are licensed under the MIT License.

Reproduction note

The exact environments, commands, outputs, limitations, and the mismatch between one historical example and the archived rerun are recorded in the reproduction record. Start there before treating any result as independently reproducible.

Validate this publication package locally with:

python3 scripts/validate_publication.py

Installable evaluation tools

The repository also provides dependency-free Python command-line tools:

python3 -m pip install ai-agent-persistence
ai-agent-persistence validate /path/to/ai-agent-persistence
ai-agent-persistence probe --runtime /path/to/ai-workshop/runtime --processes 32

The validator checks the publication's required evidence files, JSON, local links, duplicate claim identifiers, and common secret or private-path patterns. The concurrency probe launches simultaneous Saient ticks in an isolated state directory and verifies that all successful ticks are durably serialized.