Azizi Mohamed @Azizisimo

Founder of Daryl — provable memory for AI agents. Append-only, hash-chained, independently verifiable. Open source (MIT). daryl.md Istanbul, Turquie Joined July 2009

Tweets

2K
Followers

524
Following

1K
Likes

3K

Azizi Mohamed @Azizisimo

2 weeks ago

What Daryl verification does not prove. dsm verify recomputes the hash chain from the raw entries, then compares the tip to the expected integrity pin. Modify or reorder an entry, and the chain breaks. Delete the last entries cleanly, and the remaining chain still verifies — unless the recomputed tip no longer matches the pin. That is integrity against the state you expect. Not truth. An agent can append a wrong fact. The chain will still verify. It also does not, today, defeat someone who rewrites both the shard and the local pin in the same step. That needs an external witness. That is not shipped. What you get now: a trail you can replay, and a check that fails if the trail was quietly edited. DSM proves properties of the record, not reality itself. #AIAgents #DSM #OpenSource

0 0 0 11 0

View Details

Azizi Mohamed @Azizisimo

2 weeks ago

Agent memory is usually mutable. A vector store updates. A logfile gets edited. A session is patched after the fact. When something goes wrong, you are left with a story about what the agent knew — not a record you can check. Append-only means an entry is written once. Not updated. Not deleted. The chain only grows. Hash-chained means each entry’s hash is SHA-256(content + prev_hash). The next entry commits to the last one. Change a byte in the middle, and the chain no longer verifies. That is not a metaphor. dsm verify recomputes the chain from the raw entries. One altered byte, and verification fails.

0 0 0 12 0

View Details

Azizi Mohamed @Azizisimo

2 months ago

github.com/daryl-labs-ai/… - 30-45 minutes + honest feedback would really help. #OpenSource #AIagents #MCP #DeveloperTools

0 0 0 30 0

View Details

Azizi Mohamed @Azizisimo

2 months ago

What I want is simple: clone the repo, follow the Quick Start, and tell me honestly - unclear docs? harder than expected? what failed? what to simplify? If it works, great. If not, even more valuable.

1 0 1 32 0

View Details

Azizi Mohamed @Azizisimo

2 months ago

We've reached Platform M1 Release Candidate for Daryl — an open-source project, after more than a year of R&D. I'm looking for a few experienced developers to test it and tell me where it breaks.

1 0 2 3K 0

View Details

Azizi Mohamed @Azizisimo

5 months ago

The demo writes 3 sessions over 45 days. An older architecture decision ("use REST") is superseded by a newer one ("use gRPC"). Recall finds both. Temporal analysis marks the old one superseded. Context packs the relevant ones under budget. Provenance verifies the chain: integrity OK, trust verified. python demo_consumption_layer.py

0 0 0 15 0

View Details

Azizi Mohamed @Azizisimo

5 months ago

The pipeline: search_memory() — find relevant decisions across past sessions build_context() — compact them into a token-budgeted pack build_provenance() — verify chain integrity of every recalled entry Keyword scoring, temporal superseded detection, deterministic ranking. No ML. No embeddings. No network. Pure local DSM.

1 0 1 20 0

View Details

Azizi Mohamed @Azizisimo

5 months ago

Daryl v1.1 — the DSM Consumption Layer. Until now, DSM proved agent history was not tampered with. Now it also recalls that history, packages it under a token budget, and verifies its cryptographic origin. Three new modules. 77 new tests. 0 regressions.

1 0 0 35 0

View Details

Azizi Mohamed @Azizisimo

5 months ago

0 0 0 16 0

View Details

Azizi Mohamed @Azizisimo

5 months ago

This is why we call it a trust layer. Not logs. Not traces. Proof. github.com/daryl-labs-ai/… #AIAgents #TrustLayer #AI #OpenSource #CryptoVerification #AIAct #AgentSecurity

1 0 0 17 0

View Details

Azizi Mohamed @Azizisimo

5 months ago

Most AI agent systems can tell you what was recorded. Very few can tell you whether what was recorded is still true. This difference matters more than it seems. 🧵

1 0 0 18 0

View Details

Azizi Mohamed @Azizisimo

5 months ago

4/ v0.3.1 ships with: • 224/224 tests passing • Zero mandatory dependencies • Apache-2.0 licensed • External audit fixes • ADRs documenting every boundary decision Feedback, issues, stars welcome 🙏 github.com/daryl-labs-ai/… #MachineLearning #LLM #OpenSource

0 0 1 37 0

View Details

Azizi Mohamed @Azizisimo

5 months ago

3/ Integration is one line: callbacks=[DSMCallback.from_config( model_name="llama-3.1-8b", task_type="sft", )] Works out of the box with HF Trainer, TRL, and Unsloth. No patch. No fork. No config file.

1 0 0 33 0

View Details

Azizi Mohamed @Azizisimo

5 months ago

Shipped dsm-unsloth v0.3.1 today. It's a tiny bridge that gives Hugging Face training pipelines (Trainer / TRL / Unsloth) a provable, hash-chained memory of every run — metrics, evals, artifact lineage — via a single callback. github.com/daryl-labs-ai/…