ANOLISA is a toolbox, not one agent runtime
ANOLISA groups several server-side agent components under one installer and repository. cosh-ng puts an agent in a familiar shell. Token-less filters tool schemas and responses before they reach a model. AgentSight observes agent activity, while Agent Memory carries context across sessions. SkillFS, workspace checkpoints, a sandbox lifecycle tool, and Agent Sec Core cover other parts of execution. Each capability can be enabled separately, which is the right way to approach a project this broad.
The full checkout makes that breadth visible: 5,284 files, about 1,519,311 source lines, and 79.1 MB at commit 8d089b2. Our lab did not build every ANOLISA component. It entered src/agent-memory/, the Rust project identified by the supplied measurement, and evaluated that component. Claims about its clean build and tests should not be transferred to cosh-ng, Token-less, AgentSight, or the security services.
Agent Memory exposes 37 filesystem and retrieval tools over MCP
The memory component runs as one Tokio process over stdio JSON-RPC 2.0. Its README divides 37 MCP tools into file operations, structured retrieval, governance, and user-control functions. An agent can write ordinary memory files, search them, capture observations, request a token-bounded context, take snapshots, revert Git history, export data, or resume a stored task. Namespace isolation and path resolution keep operations beneath the configured memory root.
Three profiles change what the client sees. Basic favors structured retrieval, advanced favors file operations, and expert exposes file tools only. The default config is blunt about the limit: profile gating is a UX hint, not protection against a co-tenant with kernel access. A deployment that serves multiple users still needs operating-system isolation and carefully owned storage directories, even though 37 tool names can be hidden or rejected by profile.
What happened when we ran it
Our sandbox installed the Agent Memory project in 16 seconds and added 256 packages. The Rust build succeeded in 104 seconds. Cargo then completed all 692 tests in 103 seconds with 0 failures. Among these four reviews, it was the only checkout whose measured install, build, and full test step all succeeded.
The run used commit 8d089b2 in a fresh unprivileged Debian container with 3 CPUs, 12 GB of RAM, and no secrets. The repository scan found 22 CI workflow files and a tests directory, but no Dockerfile. Those facts support confidence in the measured source path. They do not measure search relevance, memory recall quality, prompt-injection detection, or the cost of an embedding provider.
Agent Memory lives inside a much larger repository, so the 1.5 million-line checkout is a real contributor cost even though the component installed only 256 packages in our run. Teams interested solely in memory should pin the component they build and watch repository layout changes. The recommended binary installer avoids a source build, while source contributors need Rust 1.85 or newer on Linux.
Hybrid search is optional and starts disabled
Without an embedding backend, the component can still store files and use keyword-oriented behavior. Vector and hybrid modes require configuration. The documented choices are OpenAI, which needs an API key, or Ollama, which needs a running local service and an embedding model. The default TOML sets the backend to none, so installing Agent Memory does not silently create semantic search. That default is sensible for privacy and makes the missing capability visible.
The index also applies time decay and can exclude cold, never-accessed files after 30 days under the shown defaults. Consolidation can extract up to 20 facts from a qualifying session log. These policies will shape what an agent recalls. Operators should test old facts, contradictions, deletions, namespace separation, and snapshots with their own conversations instead of equating a 692-test pass with correct recall for every workload.
Token-less changes context before it reaches the model
Token-less is the easiest second component to understand. It sits between an existing agent and model, dropping fields that its rules treat as redundant and replacing removed array material with retrievable markers. The README demonstrates a Claude Code adapter and gives separate statistics commands for inspecting results. Its published percentage and microsecond examples come from the project's own workload, not our lab, so they are evidence to reproduce rather than numbers to use for capacity planning.
Compression can change what the model sees even when the original tool response remains retrievable. Test it on commands where exact ordering, metadata, or empty values carry meaning. A saved token is useful only when the agent still reaches the correct decision. The adapter architecture also means support varies by client; open requests for additional tools should not be read as current compatibility.
Linux is the common ground, but component requirements differ
Agent Memory explicitly supports Linux on x86_64 and aarch64. AgentSight's code-free observation path relies on eBPF, and cosh-ng is installed in system mode with sudo in the root README. Token-less can be added independently to an existing agent. There is no single runtime recipe that proves all these paths together, and our scan found no Dockerfile to provide one shared environment.
GitHub recorded 610 stars, 161 combined issues and pull requests, and a last push on August 29, 2026. The latest release endpoint pointed to Token-less v0.7.12 from the same date, but its release body had placeholder text and an unusual comparison link. Activity is current; release notes are not equally informative across components. ANOLISA is best adopted one named component at a time, starting with Agent Memory's 692 passing tests rather than the promise of an all-in-one agent operating layer.