ResiMind

Remory is an open-source system for adding learned residual memory to large-language-model agent context compaction. It supplements an agent’s generated summary with soft memory conditioned on that summary, allowing a frozen model to continue work after earlier conversation history is discarded. During compaction, its hook reads the actor model’s hidden states and runs a memory network; during continuation, the saved soft tokens are inserted after the summary. The repository provides SGLang and Transformers deployment, model checkpoints, a Python Harness and Client API, a Codex integration, and training procedures for reconstruction and summary-conditioned residual-memory networks. It is released under the MIT License.

Reverify is an AI-agent verification toolkit for reverse engineering and related code analysis. It places deterministic tools between an agent and its claims: the model proposes hypotheses about a binary or candidate implementation, while parsers, disassemblers, pattern scanners, emulators, equivalence checks, and optional analysis engines compare them with ground truth and return evidence-backed VERIFIED, REFUTED, INCONCLUSIVE, OBSERVED, or INVALIDATED results. Its pure-Python core handles PE, ELF, and Mach-O parsing, x86/x64 and ARM/ARM64 disassembly, byte and pattern inspection, CPU micro-emulation, Protobuf/TLV dissection, and Frida hook generation; optional installations add Capstone, Unicorn, LIEF, Z3, or angr for deeper analysis. The project exposes the verification loop through a command-line interface and an MCP server that agents such as Claude Code and Cursor can call. It also records verified, observed, proved, and refuted results in a content-keyed local ledger, allowing grounded facts and negative findings to survive context resets, compaction, and new sessions. A rollover and orchestration system hands sessions off through files rather than model-written summaries, keeping verified facts separate from unverified notes. Reverify includes claim types for bytes, typed reads, instructions, patterns, emulation, behavioral or formal equivalence, imports, exports, sections, and—when angr is installed—functions, calls, references, and reachability. It is distributed from PyPI and licensed under the MIT License. The repository states that it is intended for authorized reverse engineering, malware analysis, CTF work, interoperability research, and software the user owns or is permitted to analyze.

MiniMind is an open-source tiny large-language-model project and end-to-end training tutorial developed by Jingyao Gong. Its main dense model is approximately 64M parameters, and the repository provides the model architecture, tokenizer, datasets, inference code, and training pipeline. The pipeline covers pretraining, supervised fine-tuning, hand-written LoRA, DPO, PPO, GRPO, CISPO, model distillation, tool calling, adaptive thinking, and agentic reinforcement learning. Core algorithms are implemented directly with native PyTorch rather than relying on high-level abstractions from third-party training libraries. Its agentic reinforcement-learning path performs multi-turn rollouts, executes generated tool calls, appends tool observations to the context, calculates trajectory-level rewards, and updates the policy; rollout can use local PyTorch generation or an SGLang server. MiniMind supports dense and mixture-of-experts variants, single- and multi-GPU training, YaRN-based RoPE length extrapolation, Transformers-format models, and inference through llama.cpp, vLLM, or Ollama. The repository also includes a Streamlit chat interface and a lightweight OpenAI-compatible API server with tool-call and reasoning fields. It is released under the Apache License 2.0.

Lemmalog is a Datalog engine for LLM-agent memory, distributed as a Rust crate with an MCP server, REPL, and agent skill. It treats extracted statements as provenance-tracked base facts, then uses runtime-parsed stratified Datalog to derive temporal projections, closures, contradiction candidates, relevance relationships, and aggregates. The engine supports seminaive incremental fixpoint evaluation, stratified negation with negative-cycle rejection, bi-temporal facts, confidence and provenance annotations, proof trees through why(), scoped retraction recomputation, demand-driven ask_deep queries using magic sets, and persistence of episodes, rules, and base facts while rebuilding derived relations on load. Its AgentMemory facade connects an extractor to deterministic ADD, UPDATE, NOOP, and escalation policies, and assembles budgeted contexts from distilled facts and their source episodes. The MCP server exposes the engine to agent harnesses such as Claude Code and Kimi CLI through tools for observing facts, querying, explaining proofs, retrieving context, saving state, installing rules, and running hypotheticals.

Reflexio is a learning platform for AI agents that turns user corrections, failed paths, and successful outcomes into reusable behavioral changes without retraining the underlying model. Its SDK and integration loop publish interaction outcomes, extract actionable feedback, store learned behaviors, and retrieve only relevant learnings during later inference; integrations are available through Python, REST, a CLI, and a portable coding-agent skill. Reflexio evaluates learnings against control responses and user-defined success criteria, tracking whether they solved the user's problem, required correction, or escalated to a human. It makes learnings auditable and revocable: they can be reviewed, rewritten, approved, rejected, or deleted, with rejected learnings removed from retrieval. Background processes consolidate duplicate signals and resolve conflicts or outdated lessons, while evidence from later sessions is used to revise learnings. The service supports managed, bring-your-own-key, customer-owned database, bring-your-own-cloud, and fully self-hosted deployments. The page identifies an official repository at github.com/ReflexioAI/reflexio and states that the same API can be used across deployment modes.