Desktop Apps: using the memory in Claude Desktop and the ChatGPT app
0.15.0: F7 and F34 fixed, the second formal pass counted, writes pushed only when they are writes
Wiki aligned with the code: benefits, daily use and honest maturity first
Every page audited against the source and corrected, then re-checked by a
second pass. Home leads with what the memory does, a day with it, how mature
it is and what it costs. Installation lists every kg setup item (upkeep
opt-in), updates and uninstall. Server-Management and Configuration cover
the systemd unit's journal and frozen environment. How-It-Works, Design-
Decisions and the API page match the kg mcp transport, paging, credits,
fresh tier and rebalance, push of other sessions' writes and the tool
schemas; unverifiable history is attributed. Data-and-Backup, Skills-
Reference and Visual-Editor match storage, skills and the editor, including
project live updates (F6).
0.14.0: budgets 22,000 per level and 50,000 per read
0.14.0: preload sized per client unit, paged kg_read, per-port state, maintenance credit and the one recency rule
Remove the Direction page: planning is no longer public
0.12: install through kg (uv tool install kg-memory + kg setup), kg mcp, kg editor; repo renamed to kg-memory
0.11.0: experimental Antigravity CLI, compaction reset, current-source corrections
v0.10.0: Codex CLI, a docs freshness pass, and Research and Direction pages
Every page brought up to 0.10.0: the Codex install and hook trust, file
recall, chores and their runners, the third graph level, kg_rename_node,
current endorsement rules and pass categories, the corrected .prev backup
names and the reload-not-restart recovery step, config keys and env vars
the server actually reads. Two new pages: Research (the literature read
against this system, and what changed because of it) and Direction.
Session limits and the status line: how an agent reads its own quota
New Installation section: rate_limits reaches only the status-line command's
stdin, so persisting it to ~/.claude/last-limits.json is the only channel by
which Claude can read its remaining 5h/7d budget. Covers install, the render,
and a field-semantics table — account-global percentages vs the last-renderer's
context_pct vs the updated_at freshness gate — plus the "tell Claude it exists"
step and why checkpoint-ending sessions pair with graph memory.
Home: bullet in "Also do this". Skills-Reference: kg-ops now lists the recipe,
and kg-core's recall instinct restated as layered depth-fetch to match the skill.
v0.9.24-25 docs: ambient recall + capture nudges, maintenance debt, kg-ops in skills table, server endpoints
v0.9.19 docs: compact-core preload (10K hook budget), deduped loud read, visible loading
Claude-Session: https://claude.ai/code/session_01GLF7Q6xEGRxrjXCqxXZbba
v0.9.18 docs: preloaded memory, fixed char budgets, node-centric reads, search v2, kg_useful
0.9.14: server auto-start + venv bootstrap — install flow rewritten around it
0.9.13: git-first backup story (auto-commit documented), refill single-threshold + skip semantics, security guards (Host/Origin), 5000-token default, log paths, estimator render==charge table
docs: align with v0.9.10 — simpler install, hook now bundled
Quick Install collapses to marketplace add + /plugin install + restart.
install_command.sh demoted to optional shell-symlink installer with an
idempotent legacy-hook cleanup for users upgrading from <=0.9.9.
New section: enable Claude Code plugin auto-updates for the maxim-plugins
marketplace (off by default for third-party sources, so this is the only
way to stay current without manual refreshes).
Server-Management.md: <plugin-dir> placeholders resolved with find pattern,
systemd auto-start uses cp not ln -s to survive plugin-cache version churn,
MCP config path noted as read-only (overwritten on update).
v0.9.9: update scoring formula, version bump
- Replace "recency/connectedness/richness" with "recency/connectedness + resurrection pass"
in comparison tables (richness was dropped in v0.9.9)
- Bump version to 0.9.9
v0.9.8: update kg-maintain docs — gist/notes hygiene always-run passes
docs: fix backup documentation for v0.9.6
Remove tiered backup table and fictional git auto-commit feature.
Document actual .prev crash protection and user-managed git/Borg options.
v0.9.3: three-tier compaction, skill limits, hook, corrected scoring formula
- Install: add install_command.sh as required step (not optional)
- Installation: correct skill desc limit (1,536 chars hard cap, ~5.5K total)
- Skills-Reference: rewrite hidden skill summaries, fix char count table
- How-It-Works: two-pass compaction, three-tier state table, weighted sum formula
- Configuration: fix scoring formula (product → weighted sum), add ARCHIVED_BUDGET_RATIO
- Home: bump version to 0.9.3, update compaction description, add wiki link
Update wiki for v0.9.1: compaction tuning, comparison with other memory systems
- Configuration: grace period and orphan defaults now reference constants.py
- Design-Decisions: full comparison table vs MemPalace and Claude Code Auto-Memory
- Home: "How It Compares" section with compact table and philosophical summary; version bump to 0.9.1
- How-It-Works: compaction target and grace period reference env vars, not hardcoded days
- Day/ratio values throughout generalized to reference constants.py / env vars
v0.7.2: auto-memory disable step, single CLAUDE.md recommendation, design rationale
- Installation: new step to disable built-in auto-memory; single global CLAUDE.md guidance
- Installation: troubleshooting entry for memory confusion (auto-memory + project CLAUDE.md conflicts)
- Design Decisions: why disable auto-memory and avoid project-level CLAUDE.md files
- Home: version bump to 0.7.2
Update wiki docs: centralized storage, write-through persistence, revised defaults
Initial wiki: 10 documentation pages