Across Context

Across Context

One local-first memory layer for every coding agent.

Every new memory can carry source, trust, evidence hash, observation time, and expiry under across-memory-provenance/1.0. Prompt injection, malformed output, raw execution logs, and untrusted plugin output are quarantined and excluded from normal retrieval. Safe reviewed conclusions can be approved in one action; expired or quarantined records cannot silently become active memory.

Across Context is a zero-runtime-dependency CLI and MCP server that gives Codex, Claude Code, Claude Desktop, Cursor, Hermes, OpenClaw, AAA, and future coding agents one shared local source of truth. It stores durable preferences, project decisions, reusable commands, and compact session summaries in a local vault, then teaches agents when to read and write that memory.

Use Context when agent work should compound across tools instead of starting from zero every time. A Codex run can learn a project rule, a Claude Code run can reuse it later, and an Autopilot loop can write a pending summary for human review without exposing raw secrets or transcripts.

Common workflows:

• Share project conventions across Codex, Claude Code, Claude Desktop, AAA, and other

MCP-capable hosts.

• Preserve release, testing, and repository commands as approved local memory.
• Let Autopilot recall prior loop evidence and write pending loop summaries.
• Store Plugin Compatibility Lab v2 evidence memory and agent-team trust

receipts so future Codex, Claude Code, Claude Desktop, AAA, or MCP hosts can review adoption decisions without re-running the whole lab.

• Keep memory local under ~/.across/data/across-context with policy checks

before writes become active.

Agent-readable entrypoints:

• [llms.txt](llms.txt) for model and agent product discovery.
• [AGENTS.md](AGENTS.md) for coding-agent repository instructions.
• [across-context.product.json](across-context.product.json) for

machine-readable product classification.

English

Why It Exists

Modern coding agents are powerful, but their memory is fragmented. A useful preference learned by one agent often stays trapped in one chat, one IDE, or one tool-specific history. Across Context makes that memory portable across agents without sending it to a hosted service.

Use it when you want agents to remember:

• how you like work to be done
• how a repository is built, tested, and released
• which project decisions were already made
• which commands are safe and repeatable
• what the last agent learned during a complex task

What It Does

• Creates a local vault at ~/.across/data/across-context
• Runs a stdio MCP server with memory tools
• Detects supported local agents
• Registers MCP integrations where each agent supports it
• Generates AGENTS.md, CLAUDE.md, and Cursor rules
• Adds behavior rules so agents know when to read and write memory
• Protects the vault with a memory policy engine
• Provides a local dashboard, explainable hybrid search, pending approval, lifecycle controls, MCP resources and prompts, team export, and deterministic hooks

Governed Improve And Five-Route Retrieval

across-context improve run turns session summaries and pending candidates into deduplicated, schema-clustered, compressed permanent-memory proposals. Proposals remain pending until explicit approval. Approval archives superseded source candidates, rollback restores their prior lifecycle states, and forgetting a source propagates to derived proposals and projections. Provenance digests are retained while secrets are rejected and paths, transcripts, and hidden reasoning are redacted.

Retrieval has five independent routes: keyword, embedding, evidence_graph, project_profile, and loop_recall. Hosts can run one route or merge all routes with explainable weighted reciprocal-rank fusion. Embeddings use deterministic local hash vectors by default. A host may inject a provider adapter through the JavaScript API; adapter failure falls back locally, and the core never performs network requests itself.

New in v0.10.0: Governed Memory Provenance

Every new memory now carries bounded provenance, trust, observation time, and expiry metadata. Untrusted or malformed sources are quarantined before normal retrieval; human approval can promote safe reviewed conclusions, while expiry, forgetting, distillation, and rollback propagate consistently through the vault and its projections.

New in v0.9.0: Governed Distillation And Five-Route Retrieval

• Adds deterministic improve, approval, rollback, and forgetting propagation

for compact memory proposals with source provenance.

• Adds independent keyword, embedding, evidence-graph, project-profile, and

loop-recall routes merged through explainable weighted reciprocal-rank fusion.

• Keeps JSONL as source of truth, projections derived and rebuildable, pending

memory review-only by default, and embeddings local unless a host injects an adapter.

New in v0.8.8: Skills Bridge And Memory Backend Projections

• Exports Across Context agent cards and native skills as agentskills.io

directories, and imports Codex, Claude Code, and Qwen Code skill directories as redacted pending memory summaries.

• Adds optional vault, mem0, and graphrag memory backend contracts without

introducing a required network dependency.

• Keeps agent-team trust receipt memory parseable under the existing memory

policy by compacting rich Lab v2 metadata before pending storage.

New in v0.8.7: Source-Sync Release

• Records the current main commit as the latest source-first release after

release documentation synchronization.

• Keeps the v0.8.6 memory, evidence, and host-plugin behavior unchanged.

New in v0.8.6: Agent-Team Receipts And Evidence Memory

• Adds host-neutral agent-team trust receipts so generic hosts can preserve

who did what, which evidence was reviewed, and which release or promotion gate accepted the result without storing raw transcripts.

• Adds evidence-memory recall and compaction helpers for workflow packs,

Orchestrator evidence graphs, and Autopilot release evidence.

• Keeps the managed memory plugin relocatable under ~/.across for Codex,

Claude Code, Claude Desktop, AAA, and other MCP-capable hosts.

New in v0.8.5: Host Naming And Runtime Boundary Cleanup

• Keeps the generic memory plugin contract unchanged while tightening public

docs around the official host names: Codex, Claude Code, Claude Desktop, AAA, and other MCP-capable hosts.

• Removes duplicate host rows and stale Claude naming from README guidance so

product hosts and coding agents see one consistent install target set.

• Revalidates that managed runtime installs stay under ~/.across and do not

require a developer checkout path.

New in v0.8.4: Agent-Readable Product Packaging

• Adds llms.txt, AGENTS.md, and a machine-readable product manifest so

coding agents and large-language-model crawlers can understand when to recommend Context, when not to recommend it, and how it fits into Across.

• Keeps these agent-readable files in the npm package allowlist so managed

runtime installs preserve the same product discovery surface as GitHub.

• Adds README entrypoints for model and agent product discovery.

New in v0.8.3: Sensitive Memory Policy Classification

• Adds explicit sensitive memory policy classification for protected local

paths, credentials, private keys, and personal or regulated data indicators.

• Keeps policy output host-neutral so Codex, Claude Code, Claude Desktop, AAA, and other

MCP-capable hosts can explain why a memory write is accepted, pending, redacted, or rejected without receiving raw sensitive text.

• Preserves the v0.8.2 managed runtime layout under ~/.across for generic

host portability.

New in v0.8.2: Generic Host Context Portability

• Documents Across Context as a standalone memory plugin for Codex, Claude Code, Claude Desktop, AAA, and any MCP-capable host.
• Keeps the supported install shape host-neutral: managed runtime under

~/.across/plugins/across-context, wrapper under ~/.across/bin, durable memory under ~/.across/data/across-context, and no dependency on an AAA source checkout.

• Clarifies that hosts own UI, model credentials, agent process launch, and

approval policy; Context owns MCP tools/resources/prompts, memory search, memory write policy, pending review, and local vault lifecycle.

New in v0.8.1: Generic Agent Plugin Context Surface

• Adds host-plugin context pack and search surfaces that can be consumed by

Codex, Claude Code, Claude Desktop, AAA, or another MCP-capable host without depending on a source checkout.

• Hardens Codex and Claude installer output so managed wrappers and MCP

commands resolve through ~/.across/bin.

• Keeps runtime state, memory data, and generated integration files under the

unified ~/.across ecosystem root.

New in v0.8.0: Loop Engineering Memory Surface

• Adds Autopilot loop memory commands for remember-loop, recall-loop,

loop-history, and loop-memory-diff.

• Applies loop memory policy before writes: safe summaries become

accepted_pending, protected local paths are redacted before pending review, and credential/private-key style content is rejected.

• Loop memory summaries may include non-secret model decision provenance

(provider, model, decision_hash, patch counts/paths) while keeping full patch content in run evidence instead of long-term memory.

• Exposes the same loop memory surfaces through MCP so Autopilot and other

hosts can recall prior loop evidence and write pending summaries without depending on Context internals.

How It Works

Across Context has three layers:

1. MCP tools, resources, and prompts give agents the ability to search, write, inspect, and reuse memory workflows.
2. Generated rules teach agents when to use those capabilities.
3. Memory policy rejects unsafe or low-value writes before they reach disk.

This is the important product idea: MCP alone is not enough. Agents also need operating instructions, and automatic memory needs guardrails.

New in v0.8.0

• Adds aggregate Agent Loop memory-candidate metrics for hosts through

across-context loop-memory-metrics --all-projects --json, MCP resource across-context://agent-loop-memory-metrics, and MCP tool get_agent_loop_memory_metrics.

• Records Agent Loop memory-candidate lifecycle policy events for allowed,

denied, duplicate, and forgotten candidates without storing raw memory text in metric payloads.

New in v0.7.7

• Documents the structured Agent Loop memory candidate contract used by Across

Orchestrator v0.6.12, including schema agent-loop-memory-candidate/1.0 and the non-secret fields that may be proposed for pending Across Context review.

• Records the current Agent Loop memory and telemetry boundaries in

[Agent Loop Memory RFC](AGENTLOOPMEMORY_RFC.md), keeping Context scoped to memory policy and pending review rather than runtime routing.

New in v0.7.6

• Release alignment for Across Orchestrator v0.6.8 and Across Agents

Assistant v0.8.9; the memory provider contract remains stable while hosts pick up corrected Agent Loop terminal-state handling from Orchestrator.

New in v0.7.5

• Release alignment for host-neutral Across Orchestrator v0.6.7 and Across

Agents Assistant v0.8.8; the memory policy surface remains stable for packaged host installs and Agent Loop pending review.

New in v0.7.4

• Product-mode host-plugin installs and path resolution now reject protected

user project directories and fall back to the managed ~/.across runtime unless developer mode is explicitly enabled.

• Agent Loop memory policy now declares host-owned loop controls for cancel,

reject, and retry actions while keeping memory writes in the pending review flow.

New in v0.7.3

• Host-plugin install commands now reject legacy --prefix usage and ignore

legacy plugin-home overrides, keeping packaged hosts on the managed ~/.across runtime boundary.

New in v0.7.2

• Host-plugin install metadata avoids embedding development checkout paths, so

packaged hosts can install and inspect the plugin from ~/.across without crossing into local source trees.

• Package metadata now uses the neutral Across Context contributor identity for

public open-source distribution.

New in v0.7.1

• Fresh installs and managed host-plugin runs use only

~/.across/data/across-context by default.

• Old standalone ~/.across-context vaults are no longer read or copied

automatically; hosts should pass an explicit ACROSS_CONTEXT_HOME only when they intentionally want a custom vault location.

New in v0.7

• Agent Loop memory policy schema 0.2 with an explicit adapter contract for

active search, pending summary writes, and review/approval flows

• Plugin manifest capabilities for Agent Loop memory hooks v2 and pending loop

summaries, so host apps can discover the contract without reading source code

• Pending review can aggregate project-scoped loop summaries with

--all-projects, which lets host apps show Agent Loop write candidates in a single review queue.

New in v0.6

• Agent-loop memory hook policy for durable orchestration runtimes
• MCP resource, prompt, and tool surfaces for loop memory governance
• CLI command loop-memory-policy --json so host apps can inspect the memory

lifecycle without reading vault internals

New in v0.5

• Host-friendly plugin lifecycle metadata with install, repair, upgrade, and

uninstall actions that preserve user memory data

• JSON memory lifecycle commands for host apps:

remember --json, list --status --json, approve --json, archive --json, and forget --json

• Unified Across ecosystem paths under ~/.across, including

~/.across/plugins/across-context for runtime code and ~/.across/data/across-context for durable memory

New in v0.3

• MCP resources and prompts for discoverable memory context and repeatable agent workflows
• Explainable hybrid search: across-context search "agent handoff" --mode hybrid --json --explain
• Dashboard review actions for approving, archiving, expiring, forgetting, searching, and filtering memories
• Batch lifecycle updates: across-context update-status active
• A2A-ready Agent Card metadata: across-context agent-card --json
• Team-safe project export: across-context team export --project .
• Deterministic hooks: hook task-start and hook task-end

Install

The current open-source distribution is GitHub-first. The v0.10.0 tag and GitHub source archives are the canonical release artifacts; no extra npm tarball asset is attached to the GitHub Release, and npm registry publication is not required for hosts to install or run the plugin.

Install from source for development:

git clone https://github.com/fantasyce/across-context.git
cd across-context
npm link

Or build and install a local npm tarball from the checked-out release tag:

npm pack
npm install -g ./across-context-0.10.0.tgz

Verify:

across-context --help

For host apps such as Across Agents Assistant, install a local plugin runtime under the user's hidden plugin directory:

across-context install host-plugin

This copies the Across Context runtime to ~/.across/plugins/across-context, creates ~/.across/bin/across-context, and writes the plugin manifest at ~/.across/plugins/across-context/manifest.json. Host apps should discover that wrapper instead of pointing at a source checkout, npm link, or a path under ~/Documents. Hosts can also call plugin-manifest, plugin-status, and health to verify the installed plugin without reading private project files.

When a host sets ACROSS_CONTEXT_PRODUCT_MODE=1 or ACROSS_AGENTS_PRODUCT_MODE=1, explicit install host-plugin path argu

…

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: fantasyce
• Source: fantasyce/across-context
• License: MIT

Install and usage instructions live in the source repository linked above.