Power your agent
The publisher's own command. Not a PlugMCP endpoint.
Host a copy with our keys and logs.
npx -y @astudioplus/[email protected]Claude Code has no install link for a local command. Copy the command below and run it in a terminal.
claude mcp add io.github.anvanster-tempera -- npx -y @astudioplus/tempera-mcp@0.4.29Run in a terminal.
{ "mcpServers": { "io.github.anvanster-tempera": { "command": "npx", "args": [ "-y", "@astudioplus/[email protected]" ] } }}claude_desktop_config.json
Remote URL or a command on your machine? How the two installs differ
Run on PlugMCP
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
LaunchFrom the publisher
Tempera gives Claude Code a persistent memory that learns from experience. Instead of starting fresh each session, Claude can recall past solutions, learn what works, and get smarter over time.
The Problem: Claude Code forgets everything between sessions. You solve the same problems repeatedly, and Claude can't learn from past successes or failures.
Persistent episodic memory for AI coding: capture, retrieve, brief, dream cycle, cross-project.
The Solution: Tempera captures coding sessions as "episodes", indexes them for semantic search, and uses reinforcement learning to surface the most valuable memories when relevant.
Without Tempera: With Tempera:
┌─────────────┐ ┌─────────────┐
│ Session 1 │ ──forgotten──> │ Session 1 │ ──captured──┐
└─────────────┘ └─────────────┘ │
┌─────────────┐ ┌─────────────┐ ▼
│ Session 2 │ ──forgotten──> │ Session 2 │ ◄──recalls──┤
└─────────────┘ └─────────────┘ │
┌─────────────┐ ┌─────────────┐ │
│ Session 3 │ ──forgotten──> │ Session 3 │ ◄──recalls──┘
└─────────────┘ └─────────────┘
│ │
▼ ▼
No learning Continuous improvement
┌────────────────────────────────────────────────────────────────┐
│ 1. START TASK │
│ User: "Fix the login redirect bug" │
└────────────────────────────────────────────────────────────────┘
│
▼
┌────────────────────────────────────────────────────────────────┐
│ 2. RETRIEVE MEMORIES │
│ Claude searches: "login redirect bug" │
│ Finds: "Fixed similar issue by sanitizing return URLs" │
│ + Session context: related episodes from the same task │
└────────────────────────────────────────────────────────────────┘
│
▼
┌────────────────────────────────────────────────────────────────┐
│ 3. SOLVE FASTER │
│ Claude uses past experience to solve the problem │
└────────────────────────────────────────────────────────────────┘
│
▼
┌────────────────────────────────────────────────────────────────┐
│ 4. CAPTURE SESSION │
│ Claude saves: what was done, what worked, what failed │
│ Auto-links to current session for multi-step tasks │
└────────────────────────────────────────────────────────────────┘
│
▼
┌────────────────────────────────────────────────────────────────┐
│ 5. LEARN FROM FEEDBACK │
│ User: "That memory was helpful!" │
│ → Episode utility increases │
│ → Multi-hop Bellman propagation spreads value │
│ → Session-linked episodes get boosted │
│ → Unhelpful memories fade over time │
└────────────────────────────────────────────────────────────────┘
Over time, frequently helpful knowledge rises to the top, while stale or unhelpful memories fade away — and the system itself accumulates a per-project picture of where it tends to be wrong.
Beyond the basic capture/retrieve loop, Tempera ships several higher-order surfaces. Each is opt-in but all flow through the same MCP tools — Claude can use them without any custom client code.
falsifiability score, a category, and a ValidityScope (Forever / Language / Crate / Domain / Workaround / Project). Decay rates are per-scope — universal truths never expire, project-specific conventions fade in months, workarounds expire when the underlying issue closes.verify_advance → decay → reflect → patterns → contradict → templates. Reflections turn high-signal days into prose; patterns surface themes that keep recurring; contradict probes pairs of frequently-retrieved episodes for factual disagreements; templates extract reusable step sequences from successful task clusters.tempera_brief(files, task_type?, domain?) returns pending ask-backs, the matching reasoning template, top correction categories for those files, should-have-asked triggers, and a calibration warning if the agent's track record on this kind of task is shaky.tempera_retrieve and tempera_brief both accept cross_project=true. Transferable claims (anything not project-scoped) surface across projects; Project-scoped knowledge stays bound to its codebase. Legacy captures default to non-transferable until reclassified.# Clone and build
git clone https://github.com/anvanster/tempera.git
cd tempera
cargo build --release
# Two binaries are created:
# - target/release/tempera (CLI tool)
# - target/release/tempera-mcp (MCP server for Claude Code)
cargo install tempera
On first use, Tempera downloads the BGE-Small embedding model (~128MB) for semantic search. This happens automatically and only once:
# Initialize and trigger model download
tempera init
# Output:
# 🔄 Loading embedding model (this may download the model on first run)...
# ✅ Embedding model loaded
The model is cached globally at ~/.tempera/models/ and shared across all projects.
claude mcp add tempera --scope user -- /path/to/Tempera/target/release/tempera-mcp
The --scope user flag makes it available across all your projects.
Exit and restart Claude Code to load the new MCP server.
Run /mcp in Claude Code. You should see tempera with 12 tools.
Once connected, Claude has access to these 12 tools, grouped by purpose:
tempera_session_start # is anything queued from last session?
tempera_brief(files, ...) # what does tempera know about this exact change?
tempera_retrieve(query) # then deep-dive into specific episodes
Start of session: Call tempera_session_start first to consume any pending ask-back, then tempera_brief once the file set is known.
End of task: Capture successful sessions with tempera_capture — don't wait to be asked. Capture also drafts an ask-back via Haiku if the outcome was Failure/Partial with vague intent, so the next session in this project sees the gap.
Session chaining: Episodes captured within 2 hours of each other in the same project are automatically linked. Multi-step tasks share utility and context.
Cross-project routing: New captures get a ValidityScope suggestion from the intent-extraction LLM (v0.10.3). Project-scoped claims stay bound to one codebase; Language / Crate / Domain / Workaround / Forever claims surface across projects when retrieval opts in.
Search: tempera_retrieve(query: "authentication bug")
List all: tempera_retrieve(all: true)
Show details: tempera_retrieve(query: "abc12345") # episode ID
Cross-project: tempera_retrieve(query: "sqlx ordering", scope: "cross-project")
User: Fix the login redirect vulnerability
Claude: [Calls tempera_session_start()]
No pending ask-back — proceeding.
[Calls tempera_brief(files: ["src/auth.rs", "src/router.rs"],
task_type: "bugfix", domain: "auth-middleware")]
🔴 PENDING ASK-BACK
Ask the user first: "Should auth run before or after request logging?"
📋 REASONING TEMPLATE (bugfix / auth-middleware)
url-redirect validation
1. Find every redirect target read from request input
2. Validate against allow-list or treat as untrusted
3. Add regression test with adversarial input
⚠️ TOP CORRECTION CATEGORIES for these files
- validation_skipped (3×)
- host_header_trust (2×)
Tempera says I've been wrong here before in specific ways.
Let me ask: should auth run before or after logging?
User: Before — we need an auth'd subject for log lines.
Claude: [Implements with URL allow-list, following the template steps]
Done. [Calls tempera_capture(
summary: "Fixed login redirect by validating return URL against allow-list",
outcome: "success",
validity_scope: "domain:auth-middleware"
)]
✅ Episode saved. (Cross-project — this is an auth-middleware pattern.)
User: That memory was helpful!
Claude: [Calls tempera_feedback(helpful: true, episodes: ["..."])]
Marked helpful — future retrievals for similar tasks will rank this higher.
The CLI mirrors the MCP tool surface so you can drive everything Claude does from a shell.
# Initialize Tempera
tempera init
# Capture an episode (from a session transcript or interactively)
tempera capture --session /path/to/transcript.md
# Index episodes for semantic search (or re-index)
tempera index [--reindex]
# Search memories — project-scoped by default
tempera retrieve "database connection issues"
tempera retrieve "sqlx pattern" --cross-project # v0.10.1 — pull from other projects
# Provide feedback
tempera feedback helpful --episodes abc123,def456
# Joint summary of every self-improvement signal for these files
tempera brief --files src/auth.rs,src/router.rs \
--task-type bugfix --domain auth-middleware
# Include rows from other projects (foreign rows are tagged [from <project>])
tempera brief --files src/store.rs --cross-project
# Show + clear the pending ask-back for this project (if any)
tempera session-start
# History of system-drafted clarifying questions
tempera ask-backs [--pending] [--project P]
# Log a correction the user made
tempera log-correction --category "lifetime annotations" \
--description "I assumed &str when &'a str was needed" \
--correction "use named lifetime to match trait"
# View the correction log
tempera mistakes [--top 5] # top categories
tempera mistakes --project tempera # raw list filtered
# Log a question you should have asked up front
tempera log-should-have-asked --trigger "edit auth middleware" \
--question "Which auth provider is wired up?" \
--answer "No auth — internal-only service."
# View the should-have-asked log
tempera asks --top 5
# List stored templates
tempera templates list
# Fetch a specific template
tempera templates get --task-type bugfix --domain async-rust
# Manually trigger extraction (otherwise runs in dream cycle)
tempera templates extract --max-usd 0.20
# Per-(task_type, project) verified vs declared rates
tempera calibration --project tempera --task-type bugfix
# Run the full cycle with a budget cap (default $0.50)
tempera dream --max-usd 0.50
# Run one phase, or list available phases
tempera dream --phase reflect
tempera dream --list
# Plan only — show what would happen without making LLM calls
tempera dream --dry-run
# Author yesterday's reflection (Haiku triage + Sonnet authorship if score >= 0.5)
tempera reflect [--date 2026-05-26] [--dry-run]
# Surface active factual contradictions found during dream
tempera contradict --list
# Move an episode forward in the verification chain
tempera advance-verification --episode abc123 --to tests_pass --run-id <id>
tempera advance-verification --episode abc123 --to merged --commit <sha>
tempera advance-verification --episode abc123 --to stable_no_revert --days 30
# Multi-hop Bellman propagation (run weekly)
tempera propagate --temporal
# Prune old / low-value episodes
tempera prune --older-than 90 --min-utility 0.2 --execute
# Stats + trends
tempera stats
tempera trends --project tempera --bucket weekly
# Health check + remediation
tempera doctor [--remediate --yes --target-score 90]
# Eval harness (P@5, R@5, MRR, nDCG@5 against a fixture)
tempera eval run --fixture evals/fixtures/real.jsonl --mode hybrid
# Snapshot / restore the data dir
tempera backup
tempera backup --list
tempera backup --restore 20260524T123456Z
Tempera stores everything locally in ~/.tempera/ (shared across all projects). One memory pool serves every project; the project filter is applied at query time.
~/.tempera/
├── config.toml # Configuration (all RL params configurable)
├── episodes/ # Canonical episode JSON
│ └── 2026-01-25/
│ └── <id>.json
├── jobs.sqlite # SQLite for everything indexable (see below)
├── vectors/ # Vector index (vectrust embeddings)
├── models/ # BGE-Small embedding model (~128MB)
├── reflections/ # Daily reflection markdown (v0.7.3)
├── patterns/ # Cross-day pattern pages (v0.7.4)
└── templates/ # Reasoning templates (v0.8.3)
jobs.sqlite)Everything that needs SQL lives here. Each store opens the DB on first use and runs its migration; migrations are in migrations/ and run in order.
All projects share the same pool. Cross-project routing is controlled by each episode's ValidityScope (see below) — not by separate storage.
All knobs live in ~/.tempera/config.toml. The defaults are tuned to be useful out of the box; you only need to touch this if you want to change retrieval ranking, dream-cycle behavior, or per-phase budgets.
[retrieval]
mode = "hybrid" # vector | keyword | hybrid (BM25 + vector fusion)
similarity_weight = 0.3 # Weight for semantic similarity (project mode)
utility_weight = 0.7 # Weight for learned utility (project mode)
hybrid_similarity_weight = 0.85 # RRF-normalized retrieval (hybrid mode)
hybrid_utility_weight = 0.15
recency_weight = 0.0 # Recency (0 = off, opt-in)
recency_halflife_days = 30.0
mmr_lambda = 0.7 # MMR diversity (0=diverse, 1=relevant)
min_similarity = 0.5 # Filter threshold
[bellman]
gamma = 0.9 # Discount factor for Bellman updates
alpha = 0.1 # Learning rate
propagation_threshold = 0.5 # Min similarity for propagation
max_propagation_depth = 2 # Multi-hop depth (hops)
temporal_credit_window_hours = 1
[capture]
auto_capture = true
extract_intent_llm = true # Use LLM to extract intent + claim + scope
capture_diffs = true
ask_back_on_failure = true # Draft a clarifying question on Failure/Partial captures (v0.8.5)
[dream]
default_max_usd = 0.50 # Per-cycle budget cap
stable_threshold_days = 30 # Days before Merged → StableNoRevert
triage_model = "claude-haiku-4-5-20251001"
reflect_model = "claude-sonnet-4-6"
# Patterns phase
patterns_lookback_days = 30
patterns_min_evidence = 3
patterns_cluster_threshold = 0.75
# Contradict phase
contradict_top_n = 50
contradict_min_similarity = 0.6
contradict_max_similarity = 0.95
contradict_max_pairs = 30
contradict_min_confidence = 0.7
# Templates phase (v0.8.3)
templates_min_evidence = 3
templates_min_verification_weight = 0.30 # 0.30 = Untested (lenient); 0.60 = Merged
[storage]
max_age_days = 180 # Max episode age for pruning
min_utility_threshold = 0.05 # Min utility to keep
min_retrievals = 2 # Min retrievals before pruning allowed
consolidation_threshold = 0.85 # BKM merge threshold
cluster_threshold = 0.85
stale_age_days = 30
stale_utility_threshold = 0.2
Decay rates are scope-aware (per the ValidityScope on each episode's claim):
Value from helpful episodes spreads through the similarity graph in multiple hops:
Hop 0: Source episodes (high helpfulness, ≥2 retrievals)
│
▼ γ¹ discount
Hop 1: Similar episodes updated
│
▼ γ² discount
Hop 2: Episodes similar to hop-1 updated
│
▼ Converges when no updates occur
Episodes captured within 2 hours of each other in the same project are automatically linked:
Session abc123:
├── Episode 1: "Investigated auth bug" (debug)
├── Episode 2: "Found root cause in token validation" (research)
└── Episode 3: "Fixed token expiry check" (bugfix, success)
↓
Temporal credit flows back to episodes 1 & 2
Session-linked propagation boosts all 3
A budgeted background pipeline that runs nightly (or on demand). Each phase shares a CostBudget; free phases ignore it, paid phases check try_spend() before each LLM call.
verify_advance → decay → reflect → patterns → contradict → templates
(free) (free) (Sonnet) (Sonnet) (Haiku) (Sonnet)
↓ ↓ ↓ ↓
reflections/ patterns/ contradictions templates/
Merged to StableNoRevert after stable_threshold_days.(task_type, domain), extracts reusable step sequences via Sonnet.Worst case per full cycle: roughly $0.50 with default settings.
Retrieval ranking combines three signals with normalized weights:
score = (sim_w × similarity + util_w × utility + rec_w × recency) / (sim_w + util_w + rec_w)
Default in hybrid mode: 85% similarity (RRF-normalized over vector + BM25), 15% utility, 0% recency. The VerificationState of each episode multiplies into salience — well-verified successes weigh more.
Every claim carries a ValidityScope that determines:
is_transferable() returns true for everything except Project { name }. The retrieve and brief surfaces use this to decide what surfaces when the agent opts into cross_project=true.Legacy episodes captured before v0.6.4 don't have a scope set, so they stay project-bound by default. New captures (v0.10.3+) get a scope suggested automatically by the intent-extraction LLM call — using a colon-encoded format like language:rust, crate:[email protected], domain:async-rust, workaround:repo#123, or project. The default when in doubt is project, keeping the system conservative.
Run periodically to keep memory healthy:
# Nightly: dream cycle (verify_advance + decay + reflect + patterns + contradict + templates)
tempera dream --max-usd 0.50
# Weekly: Propagate utility values (multi-hop with convergence)
tempera propagate --temporal
# Monthly: Clean up old/useless episodes
tempera prune --older-than 90 --min-utility 0.2 --execute
# As needed: Check trends
tempera trends
# As needed: Review and consolidate
# (via MCP) tempera_review(action: "consolidate")
# As needed: health check + auto-remediate
tempera doctor --remediate --yes
The dream cycle is the load-bearing piece for long-running memory hygiene. It uses Haiku for cheap gating and Sonnet for authorship — the default $0.50 cap is the worst case across every phase.
ls /path/to/tempera-mcpcat ~/.claude.json/mcp to verifyThe BGE-Small model (~128MB) downloads on first use from HuggingFace. This requires internet access. After download, the model is cached at ~/.tempera/models/ and works offline.
Run tempera index to create/update the vector database.
If behind a firewall or proxy, ensure access to huggingface.co. The model files are downloaded via HTTPS.
tempera_brief returns "nothing to surface"This is normal early on — the brief joins against signal data (mistakes, asks, templates, calibration) that accrues over time. Specifically:
(task_type, domain) pair (templates accrue during the dream cycle).Fall back to tempera_retrieve for episode-level recall.
tempera retrieve --cross-project finds nothingEpisodes captured before v0.6.4 don't have a ValidityScope set, and v0.10's cross-project filter treats unscoped claims as project-bound (conservative default). Either (a) capture new episodes with v0.10.3+, which auto-suggests a scope, or (b) manually classify legacy episodes via the MCP validity_scope parameter on capture.
Apache 2.0
Contributions welcome! Please open an issue or PR.
© 2026 MCP Directory.
Persistent episodic memory for AI coding: capture, retrieve, brief, dream cycle, cross-project.
{ "mcpServers": { "io.github.anvanster-tempera": { "command": "npx", "args": [ "-y", "@astudioplus/[email protected]" ] } }}Tools
No tool list yet. Remote servers get one after the first health check.
FAQ
Copy the stdio config for Cursor, Claude Desktop, Claude Code, or VS Code on this page. It uses the publisher's install, not a PlugMCP proxy.
No required environment variable is listed on this record.
Tempera runs locally over stdio.
Last checked 2026-11-09. Persistent episodic memory for AI coding: capture, retrieve, brief, dream cycle, cross-project.